SEPTEMBER 2022 Sponsored by

How Machine Learning,

INSIDE:

How Machine Learning, AI & Deep

Learning Improve Cybersecurity >>

AI & Deep Learning How Can AI Help in the Fight Against

Ransomware? >>

Criminals Use Deepfake Videos to

Improve Cybersecurity
Interview for Remote Work >>

EU Debates AI Act to Protect Human

Rights, Define High-Risk Uses >>

Machine intelligence is influencing all aspects of cybersecurity. Here’s what Deep Instinct Perspectives: Harnessing
security teams need to consider to make AI work in their organizations. the Power of Deep Learning to Speed Up
Threat Detection >>

Next
FEATURE
How Machine Learning, AI & Deep Learning
Improve Cybersecurity
Machine intelligence is influencing all aspects of cybersecurity. Here’s what security teams need to consider
to make AI work in their organizations.
By George Hulme, Contributing Writer, Dark Reading
E
nterprise security teams are strained. They’re
striving to defend ever-more entryways as at-
tackers’ sophistication and motivation grow ex-
ponentially. At the same time, enterprise environments
are becoming increasingly complex, with more servers,
endpoints, cloud systems, and applications coming on-
line daily.
To even have a chance at reducing risks, enterprises
have begun deploying security solutions at a breakneck
pace, installing (by some estimates) an average of 76
security tools and spending upward of 14% of their an-
nual IT budgets on security. The situation is untenable.
“Enterprise security teams are just deluged with threats
and priorities,” says Scott Crawford, an analyst at 451
Research. “A lot of the tools that they are turning to are
going unused because they don’t have the staff to man-
age them.”
September 2022 2
That staffing shortage is unlikely to let up any time
soon. With an estimated 2.7 million gap in skilled cy-
bersecurity professionals needed to fill current demand,
one should expect enterprise security teams to remain
strained for some time to come. However, the rise of ar-
tificial intelligence (AI) and machine learning (ML) to aug-
ment human security efforts offers a glimmer of hope at
alleviating the pain.
How Can AI/ML Help?
Advancements in AI technology and ML algorithms prom-
ise to transform enterprise security by giving IT secu-
rity teams enhanced tools to better detect and respond
to attacks faster than before. Before security teams can
use these tools, they need to know what is possible with
different types of learning and how AI can best address
today’s complex cybersecurity challenges. Although it is
Previous Next
HOW MACHINE LEARNING, AI & DEEP LEARNING IMPROVE CYBERSECURITY
common to hear these terms used interchangeably, se-
curity teams need to understand the differences between
machine learning and deep learning, and how to evaluate
and choose the right methods for success.
If AI/ML can do anything, it can help human defenders
become more efficient. “AI systems can support human
defenders by saving time, cutting costs, and reducing po-
tential for human error,” says Jason Rader, vice president
and chief information security officer at services provid-
er Insight. “AI’s ability to speed up processes that would
take hours — or perhaps days or months — for humans
to complete is substantial.”
According to Ed Bowen, AI leader and managing direc-
tor at Deloitte’s AI Center of Excellence, AI/ML can, using
techniques such as reinforcement learning, help improve
functions from vulnerability testing through identity and
access management. “We are also seeing exciting devel-
opments in the ability of AI models and tools to increase
the efficacy of cyber-threat detection and response, and
to expedite cyber operators’ reaction time to the greatest
threats,” says Bowen. AI/ML’s ability to reduce noise in
alerts and data should enable human defenders to spend
more time focused on high-value security efforts.
DJ Singh, cybersecurity architect at Capgemini Amer-
icas, echoes Bowen’s observations and says customers
are seeking ways to implement AI-based security that will
analyze event data using ML models that identify attack
patterns. “These specialized tools help narrow down the
events to those most likely to be actual risks, resulting
in faster, near-real-time detection of adversarial activity,”
Singh says.
Ultimately, AI/ML proponents say these technologies
significantly reduce the time to detect breaches, en-
hance cybersecurity decision-making through boosted
data analysis, improve overall security defenses through
enhanced attack and user behavioral analysis, and even
increase effective automation. “Direct business benefits
can include reduced analyst workload, decreased risk
Experts advise that security teams should not take the homegrown AI/ML path
unless they are really committed to having a large supporting organization with
a lot of committed resources.
exposure through better coverage, and significantly low-
er security costs. AI/ML-based security tools also enable
better breach recovery times with an automated response
capability,” says Singh.
For those reasons, analysts expect the AI/ML market to
continue to grow more rapidly than the overall enterprise
technology market. Market intelligence firm Meticulous
Research expects the cybersecurity market to continue
to grow at 24% annually through 2029, when it is expect-
ed to reach $66 billion.
September 2022 3
How AI/ML Is Used in Cybersecurity
How are enterprises predominantly deploying AI/ML?
Are they aggregating their own data and assigning
their homegrown algorithms to assess that data, or
is the intelligence embedded within the products and
services they are using? Experts say most enterprises
are turning to AI/ML within the tool sets they acquire
and deploy.
“I do see the more advanced security teams applying
data science and machine learning to their own data-
sets,” says Allie Mellen, an analyst at Forrester Research.
“This is a very hard thing to do because of the
specialized staffing required. You need data scientists
and people who understand the threats that are in the
environment. You need to understand the environment
yourself. It gets very costly and very resource inten-
sive very quickly,” she adds. Mellen and other experts
advise that usually security teams should not take the
homegrown AI/ML path unless they are really commit-
ted to having a large supporting organization with a lot
of committed resources.
Previous Next
HOW MACHINE LEARNING, AI & DEEP LEARNING IMPROVE CYBERSECURITY
451’s Crawford agrees: “Building one’s own AI/ML is
something that’s not very common and something we
see more among the security one-percenters.”
Increasingly, security vendors and service providers
are incorporating AI/ML capabilities within their tech-
nologies. “Security vendors have significantly more re-
sources to provide to those with data science skills,
such as better salaries, better benefits, and the ability
to provide more opportunities for career advancements.
Vendors can bring the resources needed to succeed in
the role, and that’s not always an option for security
Actual AIs actively interact with environments and attempt to understand their
environments enough to make useful and accurate decisions much like a
human would.
teams within the enterprise,” Mellen says.
“We very much need the human element in cyberse-
curity to make it successful,” Mellen says. “Most of the
use cases that I’ve seen around machine learning are
on detection, particularly anomaly detection, and find-
ing patterns and being able to identify when an activity
is happening that shouldn’t be happening.”
Security teams must understand the various types of
AI/ML available, how to vet vendor AI/ML claims, and
how to deploy AI/ML technologies effectively.
Assessing the Vendor’s AI/ML Claims
When CISOs and others evaluate the marketing claims
of security vendors boasting their AI/ML capabilities, it’s
tough to tell what’s real and what’s not. “That’s really the
rub there,” says Crawford. “When someone is putting for-
ward an approach to AI/ML as part of their value propo-
sition, it becomes difficult for the person who isn’t a data
scientist or isn’t familiar with the techniques to know if the
claims are real.”
“Among one of the biggest bait-and-switches in AI
vending is the sale of rules-based engines in place of AI/
ML programs,” says Adam Kamor, co-founder and head
of engineering at synthetic data provider Tonic.ai. “AI
learns from data while rules-based engines do not. This
would allow an AI program to handle unexpected out-
comes because it has seen the unexpected event in past
data. An AI model is less likely to overlook or miss small
patterns in data that someone engineering a rules-based
engine may [miss].”
To avoid such bait-and-switches and promises vendors
cannot deliver, understanding the terms is crucial.
September 2022 4
What’s the difference between AI and ML? That de-
pends on whom you ask, but many contend that machine
learning isn’t an AI because it doesn’t actively interact
with the environment. Rather, ML collects and categoriz-
es data, and then makes observations and predictions on
the data it ingested using mathematical modeling. Actual
AIs, however, actively interact with environments and at-
tempt to understand their environments enough to make
useful and accurate decisions much like a human would.
Machine learning is a subset of the study and implemen-
tation of artificial intelligence.
There are also different types of learning, specifically su-
pervised and unsupervised learning. Supervised learning
uses data inputs that are already labeled that train ma-
chine learning systems. With unsupervised learning, the
data inputs are not previously labeled, and the machine
intelligence finds patterns and useful information within
the data without human help. Most enterprise security
teams need AI/ML systems to identify conditions humans
cannot, or provide accurate analysis to surprising (unex-
pected) situations and improve analysis over time.
Although today’s security challenges and enterprise
environments are more sophisticated than ever, some
use cases are relatively straightforward to predict with
enough historical data, such as many types of fraudu-
lent financial transactions and end-user behavioral anal-
ysis. Also, AI/ML is useful for analyzing vulnerability
Previous Next
HOW MACHINE LEARNING, AI & DEEP LEARNING IMPROVE CYBERSECURITY
assessments much more rapidly and accurately than
humans or spotting flaws in application code that de-
velopers miss. Conversely, some attacks can be quite
complex, and events and techniques an attacker uses to
move laterally may not be as one would expect. Here,
unsupervised learning systems can identify causations
human analysts may miss, and ideally, this would occur
before attacks are successful.
To help discern if tools are AI or simple rule-based anal-
ysis, Kamor suggests CISOs ask their vendors how their
product or service handles unexpected outcomes and
whether accuracy improves without human involvement.
“One can automate AI models to learn from new data as
it comes in, which cannot be done with rule-based sys-
tems. At the end of the day, an AI model is making pre-
dictions based on data it has seen previously. The gain
here comes from the automation of the AI model learning
from new data that comes in to make predictions on even
newer data,” Kamor advises.
That means, ultimately, enterprise teams must test ven-
dor claims and the capabilities of their AI/ML systems.
Typically, explains Kamor, that type of quality-assurance
testing is conducted in production environments. How-
ever, before being deployed into production, the AI/ML
needs to be tested through training and validation phases
that rely on synthetic and non-production data. “The first
stage of testing requires an evaluation of training data
after it’s been put through a training environment. During
this testing, engineers need to watch for biases that will
skew results, produce patterns, and create false correla-
tions,” Kamor advises.
Capgemini Americas’ Singh says each organization is
unique and will be quite different from any test or training
data provided by a vendor to teach the AI/ML models.
Singh advises enterprises to find a methodical approach
that will help them qualify the AI/ML vendor’s offering
based on measurable business benefits.
Before being deployed into production, the AI/ML needs to be tested
through training and validation phases that rely on synthetic and non-
production data.
Singh recommends that any assessment address the
following questions:
•  Does the AI/ML have out-of-the-box integrations with
data sources — for instance, identity, endpoint, and
network events?
•  Does the AI/ML have out-of-the-box integrations
with security information and event management and
incident response tooling?
•  How granular is the application programming interface?
Does it allow additional data modeling using tools such
as GraphQL?
September 2022 5
•  Can the AI model be explained as to why it has made a
certain decision? Or is the AI/ML model a “black box”?
•  Can the AI model provide enough details to help with
forensics?
Many of the experts we interviewed suggest security
teams would do best to ignore all vendor claims regarding
their AI/ML capabilities. “In most cases, security teams
won’t be vetting various types of AI/ML — they will be
seeking improved outcomes,” says 451’s Crawford. That
means they’ll focus on whether they see a noticeable im-
provement in the signal-to-noise ratio within the data so
that human analysts get more actionable insights.
“That’s a practical measure that people can use, and
to the extent that tools deliver that, then practitioners will
see benefit, regardless of the underlying technologies. It’s
really about the efficacy and effectiveness of the technol-
ogy,” he says.
Sohrob Kazerounian, AI research lead at Vectra AI,
agrees. “The bottom line is the bottom line,” Kazerounian
says. “If AI/ML is not materially improving some business
objective, then it’s a distraction, not a solution. AI/ML
Previous Next
HOW MACHINE LEARNING, AI & DEEP LEARNING IMPROVE CYBERSECURITY
vendors need to have clear outcome-based value prop-
ositions, and CISOs need to hold those vendors’ feet to
the fire to prove them.”
Kazerounian advises that CISOs focus their testing
efforts on outcomes. “Tests should positively validate
that successful outcomes are material improvements
over whatever they’re replacing. While it isn’t always
feasible to test in a production environment, some kind
of live-fire exercise will inevitably provide the best re-
sults,” he says.
The security team should have clear plans about how the AI/ML deployment
will benefit the entire organization, as well as the security team, and plan for the
AI/ML’s ongoing support.
Implementing AI/ML in the Real World
Just as AI/ML vendors claim assessment tests need
clear outcomes, so do the pilots and deployments. Ton-
ic.ai’s Kamor says, “Every AI program needs to have a
clear directive, and too often, enterprises dive in without
a clear vision.”
Capgemini’s Singh agrees that any AI/ML implementa-
tion should have a clearly defined success criteria — and
detailed in business terms, not technical terms — such as
reduction in manual effort, detection of advanced attacks,
reduction in risk, cost avoidance due to early detection,
fidelity of event data, or savings in security information
and event management (SIEM) storage costs.
And before any implementation, the security team
should have clear plans about how the AI/ML deployment
will benefit the entire organization, as well as the securi-
ty team, and plan for the AI/ML’s ongoing support. “The
organization should create a detailed execution plan, un-
derstand what is achievable, and have the budget to con-
tinue the program — not just budget for the deployment,”
Crawford says.
“Be prepared to see what needs adjustment in the mod-
el. It’s always exciting for innovators and developers to
come up with new approaches, new models, but where
they really demonstrate their value is in their proper tun-
ing to a given use case and over time,” Crawford says.
For detection, that would be tuning to improve attack
identification and response, but it also applies to other
areas, such as improving entitlement management, vul-
nerability assessments, and even attack simulations. In-
sight’s Rader advises to also look at the addition of AI/ML
strategically, and not tactically.
September 2022 6
“Does it integrate into my current platform? Is there an
API integration? Does this work across my on-prem and
cloud workloads? What are the security implications of
the product itself? What are the automation capabilities?
It’s very probable that the AI/ML features of that shiny,
new product are abstracted from the operators on the se-
curity team, so the usability and integration features into
our current security operations, and the time to value are
key factors,” Rader says.
“Enterprise security teams must also remember that AI/
ML systems rely on consistency of data. At times, se-
curity vendors make changes that effect the granularity
of event data. AI/ML is not a ‘set it and forget it’ tool. It
requires re-testing and its own qualifications when data
changes,” Capgemini’s Singh says.
Singh says that new tools and conditions can impact
an AI/ML’s learning behavior, and to mitigate the risks as-
sociated with such changes, teams should consider im-
plementing AI/ML with a nod to the modern developer’s
playbook: Implement phases of small changes and test
and release to a staging environment. “This incremental
approach not only helps the AI/ML model to learn the
new data source, but also helps fine-tune the data source
one at a time. And it creates a good test environment,
including tooling to simulate complex attacks, as well as
test data that is a representative sample of a real-life pro-
duction environment,” he adds.
Previous Next
HOW MACHINE LEARNING, AI & DEEP LEARNING IMPROVE CYBERSECURITY
Experts offer other advice for enterprises aiming to
avoid missteps. “Organizations that don’t have a ma-
ture model development life cycle and an AI risk and
governance program can run into AI implementation
challenges at various points,” Deloitte’s Bowen says.
Singh points out areas in which organizations may
face challenges when implementing AI without a ma-
ture model:
•  Insufficient data profiling.
•  Not assembling representative datasets for training
models.
•  Not testing multiple model architectures to determine
the best performance for a given problem.
•  Using inappropriate performance metrics (for example,
trying to measure the “accuracy” of the model when
the dataset is highly imbalanced).
•  Skipping the best practice of sufficient quality control
and peer review throughout the model development
life cycle.
Going forward, as Forrester’s Mellen points out, the in-
vestments made by security services providers today in
AI/ML will eventually make their way into the products
enterprise security teams use. That’s because cyberse-
curity vendors, especially those with a services arm, are
increasingly using AI/ML to determine ways to improve
the effectiveness and the quality of their analyst response
by analyzing the data and determining what data should
be brought to the attention of analysts so that they can
perform their jobs better.
“We’ve yet to see that really cascade down to the prod-
uct side of things. I think it’s very interesting because be-
ing able to provide things like recommendations to secu-
rity analysts on what responses they should take could
be a very compelling use case that we haven’t explored
yet because we’ve been so caught up in the detection
aspects,” Mellen says.
Such a future should strike most security professionals
as good news, as they need as much help — human or
machine — as they can get when it comes to making
sense of their environments, attack techniques, and how
to make the most effective responses possible. And that
help can’t get here soon enough.
About the Author: For more than 20 years, George V. Hulme has
written about business, technology, and cybersecurity topics. He
currently focuses on cybersecurity and digital innovation. Previ-
ously, he was senior editor at InformationWeek magazine, and has
freelanced for many trade and general-interest publications.
September 2022 7
Previous Next
ANALYSIS
How Can AI Help in the Fight Against
Ransomware?
Less than a quarter of organizations believe they are fully prepared for a ransomware
attack, threatening data privacy.
By Maxine Holt, Research Director, Omdia
I
ndividuals are increasingly aware of the importance of
data privacy, and governments continue to implement
and tighten associated regulations. How successful-
ly are organizations dealing with data privacy? It varies
wildly; there are all-too-frequent reports of data privacy
failures, often associated with ransomware. A Dark Read-
ing poll that ended in December 2021 found that less than
a quarter of organizations believe they are fully prepared for
a ransomware attack, leaving the remaining three-quarters
highly susceptible, which in turn threatens data privacy.
Ransomware will continue to be a hugely successful
method of attack that organizations must defend against,
with data privacy regulations a significant part of the
equation. Focusing on the information life cycle (create,
process, store, transmit, destroy) will help organizations
understand what data requires protection and where it
resides. Furthermore, classifying data appropriately is
September 2022 8
important as all data is not equal: Some data will require
strong protection, and other data will not. By understand-
ing these nuances, organizations can begin exploring
more advanced approaches to ransomware as with the
use of artificial intelligence (AI) to see unseen patterns in
the data that may point to a potential incursion or threat.
Enter Deep Learning
Attackers using malware can block access to data and/
or systems, encrypt and lock data, or even move compa-
ny data off-site. Attacks that take place over a keyboard
can be particularly difficult to detect and mitigate as they
can dwell over time, appearing innocuous at first as at-
tackers may use trusted routes of ingress as they move
laterally through a target network. AI techniques such as
unsupervised deep learning (DL) can help organizations
understand attack targets and vectors by encouraging ob-
Previous Next
HOW CAN AI HELP IN THE FIGHT AGAINST RANSOMWARE?
servability across the data life cycle. If an organization can
detect the wake of activity created by a potential wrong-
doer, it stands a good chance of blocking or diverting an
incursion before systems can be locked or data encrypted.
Here, AI offers many helpful tools that can help compa-
nies deal with malware. Statistical and mathematical ma-
chine learning (ML) algorithms like “k-nearest neighbor”
and “decision trees” can identify malware payloads and
known attack patterns, for example. Where AI really steps
into the spotlight, however, is with DL neural networks.
Unlike statistical and mathematical ML technologies that
use known rules (e.g., “this is or is not a piece of malware”)
to identify a potential attack, DL technologies can actu-
ally deduce the rules themselves. Popular DL algorithms
— including convolutional neural networks (CNNs), recur-
rent neural networks (RNNs), and long short-term mem-
ory (LSTM) — can parse huge amounts of disparate data
to build an understanding of the patterns in that data,
patterns that may turn out to represent an attack.
Data Governance is Key
IT and security practitioners considering investing in AI as
a means of fighting ransomware must first build an under- a regular speaker at events and writes a monthly Computer Weekly
standing of their entire data landscape as it pertains to data article covering various aspects of information security.
security and privacy. This means building solid metadata
defining ownership, access, privacy exposure, locality, and
so on. On top of this, the organization must establish a set
of governance requirements that span the full information
life cycle (create, process, store, transmit, destroy). Fortu-
nately, both within and beyond the confines of the security
industry, technology providers are presently laser-focused
on helping companies build a consistent view of company
operational, system, and analytical data using the concept
of a data fabric.
Over time, Omdia expects these metadata efforts to
more closely align between security and business practic-
es. At that time, companies will likely provision an AI-ca-
pable malware tool in the same way they provision any
cloud-native service, by specifying data sources and flip-
ping the “on” switch. Until then, organizations without an
existing investment in a data fabric may find themselves
somewhat handicapped without the ability to “observe”
the entirety of the system of resources they’re seeking to
protect. In other words, fighting malware, just like fighting
data privacy risks, demands a high degree of data literacy,
domain expertise, and governance.
About the Author: Maxine Holt leads Omdia’s cybersecurity
research, developing a comprehensive research program to
support vendor, service provider, and enterprise clients. Maxine is
September 2022 9
Previous Next
ANALYSIS
Criminals Use Deepfake Videos to Interview
for Remote Work
The latest evolution in social engineering could put fraudsters in a position to commit insider threats.
By Ericka Chickowski, Contributing Writer, Dark Reading
S
ecurity experts are on the alert for the next evolution
of social engineering in business settings: deepfake
employment interviews. The latest trend offers a
glimpse into the future arsenal of criminals who use con-
vincing, faked personae against business users to steal
data and commit fraud.
The concern comes following a recent advisory from the
FBI Internet Crime Complaint Center (IC3), which warned of
increased activity from fraudsters trying to game the online
interview process for remote-work positions. The advisory
said that criminals are using a combination of deepfake vid-
eos and stolen personal data to misrepresent themselves
and gain employment in a range of work-from-home po-
sitions that include information technology, computer pro-
gramming, database maintenance, and software-related
job functions.
Federal law-enforcement officials said in the advisory that
they’ve received a rash of complaints from businesses.
September 2022 10
“In these interviews, the actions and lip movement of
the person seen interviewed on-camera do not complete-
ly coordinate with the audio of the person speaking,”
the advisory said. “At times, actions such as coughing,
sneezing, or other auditory actions are not aligned with
what is presented visually.”
The complaints also noted that criminals were using sto-
len personally identifiable information (PII) in conjunction
with these fake videos to better impersonate applicants,
with later background checks digging up discrepancies
between the individual who interviewed and the identity
presented in the application.
Potential Motives of Deepfake Attacks
While the advisory didn’t specify the motives for these at-
tacks, it did note that the positions applied for by these
fraudsters were ones with some level of corporate access
to sensitive data or systems.
Previous Next
CRIMINALS USE DEEPFAKE VIDEOS TO INTERVIEW FOR REMOTE WORK
Thus, security experts believe one of the most obvious
goals in deepfaking one’s way through a remote inter-
view is to get a criminal into a position to infiltrate an
organization for anything from corporate espionage to
common theft.
“Notably, some reported positions include access to
customer PII, financial data, corporate IT databases and/
or proprietary information,” the advisory said.
“A fraudster that hooks a remote job takes several giant
steps toward stealing the organization’s data crown jew-
els or locking them up for ransomware,” says Gil Dabah,
co-founder and CEO of Piiano. “Now they are an insider
threat and much harder to detect.”
Additionally, short-term impersonation might also be a
way for applicants with a “tainted personal profile” to get
past security checks, says DJ Sampath, co-founder and
CEO of Armorblox.
“These deepfake profiles are set up to bypass the
checks and balances to get through the company’s re-
cruitment policy,” he says.
There’s potential that in addition to getting access for
stealing information, foreign actors could be attempting
to deepfake their way into US firms to fund other hack-
ing enterprises.
“This FBI security warning is one of many that have been
reported by federal agencies in the past several months.
Recently, the US Treasury, State Department, and FBI
released an official warning indicating that companies
must be cautious of North Korean IT workers pretending
to be freelance contractors to infiltrate companies and
collect revenue for their country,” explains Stuart Wells,
CTO of Jumio. “Organizations that unknowingly pay North
Korean hackers potentially face legal consequences and
violate government sanctions.”
What This Means for CISOs
A lot of the deepfake warnings of the past few years have
been primarily around political or social issues. However,
this latest evolution in the use of synthetic media by crimi-
nals points to the increasing relevance of deepfake detec-
tion in business settings.
“I think this is a valid concern,” says Dr. Amit Roy-Chow-
dhury, professor of electrical and computer engineering
at University of California at Riverside. “Doing a deep-
fake video for the duration of a meeting is challenging
and relatively easy to detect. However, small companies
may not have the technology to be able to do this detec-
tion and hence may be fooled by the deepfake videos.
Deepfakes, especially images, can be very convincing
and if paired with personal data can be used to create
workplace fraud.”
Sampath warns that one of the most disconcerting
parts of this attack is the use of stolen PII to help with
the impersonation.
September 2022 11
Previous Next
CRIMINALS USE DEEPFAKE VIDEOS TO INTERVIEW FOR REMOTE WORK
“As the prevalence of the Dark Net with compromised
credentials continues to grow, we should expect these
malicious threats to continue in scale,” he says. “CISOs
have to go the extra mile to upgrade their security pos-
ture when it comes to background checks in recruiting.
Very often these processes are outsourced, and a tighter
procedure is warranted to mitigate these risks.”
Future Deepfake Concerns
Prior to this, the most public examples of criminal use of
deepfakes in corporate settings have been as a tool to
support business email compromise (BEC) attacks. For
example, in 2019 an attacker used deepfake software to
impersonate the voice of a German company’s CEO to
convince another executive at the company to urgently
send a wire transfer of $243,000 in support of a made-
up business emergency. More dramatically, last fall a
criminal used deepfake audio and forged email to con-
vince an employee of a United Arab Emirates company
to transfer $35 million to an account owned by the bad
guys, tricking the victim into thinking it was in support of
a company acquisition.
According to Matthew Canham, CEO of Beyond Layer
7 and a faculty member at George Mason University, at-
tackers are increasingly going to use deepfake technol-
ogy as a creative tool in their arsenals to help make their
social engineering attempts more effective.
September 2022 12
“Synthetic media like deepfakes is going to just take
social engineering to another level,” says Canham, who
last year at Black Hat presented research on counter-
measures to combat deepfake technology.
The good news is that researchers like Canham and
Roy-Chowdhury are making headway on coming up with
detection and countermeasures for deepfakes. In May,
Roy-Chowdhury’s team developed a framework for de-
tecting manipulated facial expressions in deepfaked vid-
eos with unprecedented levels of accuracy.
He believes that new methods of detection like this
can be put into use relatively quickly by the cybersecu-
rity community.
“I think they can be operationalized in the short term —
one or two years — with collaboration with professional
software development that can take the research to the
software product phase,” he says.
About the Author: Ericka Chickowski specializes in coverage of
information technology and business innovation. She has focused
on information security for the better part of a decade and reg-
ularly writes about the security industry as a contributor to Dark
Reading.
Previous Next
ANALYSIS

EU Debates AI Act to Protect Human Rights,

Define High-Risk Uses
The commission argues that legislative action is needed to ensure a well-functioning market for AI systems
that balances benefits and risks.
By Nathan Eddy, Contributing Writer, Dark Reading

T
he European Commission (EC) is currently debating new rules and actions
for trust and accountability in artificial intelligence (AI) technology through a
legal framework called the EU AI Act. Its aim is to promote the development
and uptake of AI while addressing potential risks some AI systems can pose to safe-
ty and fundamental rights.
While most AI systems will pose low to no risk, the EU says, some create dangers
that must be addressed. For example, the opacity of many algorithms may create
uncertainty and hamper effective enforcement of existing safety and rights laws.
The EC argues that legislative action is needed to ensure a well-functioning inter-
nal market for AI systems where both benefits and risks are adequately addressed.
“The EU AI Act aims to be a human centric legal-ethical framework that intends to
safeguard and protect human rights and fundamental freedoms from violations of
these rights and freedoms by algorithms and smart machines,” says Mauritz Kop,
Transatlantic Technology Law Forum Fellow at Stanford Law School and strategic
intellectual property lawyer at AIRecht.
The right to know whether you are dealing with a human or a machine — which is
becoming increasingly more difficult as AI becomes more sophisticated — is part of

September 2022 13
Previous Next
EU DEBATES AI ACT TO PROTECT HUMAN RIGHTS, DEFINE HIGH-RISK USES
that vision, he explains.
Kop notes that AI is now mostly unregulated, except
for a few sector-specific rules. The act aims to address
the legal gaps and loopholes by introducing a product
safety regime for AI.
“The risks are too high for nonbinding self-regulation
by companies alone,” he says.
Effects on AI Innovation
Kop admits that regulatory conformity and legal com-
pliance will be a burden, especially for early AI startups
developing high-risk AI systems. Empirical research
that shows the GDPR — while preserving privacy and
data protection and data security — had a negative ef-
fect on innovation, he notes.
Risk classification for AI is based on the intended
purpose of the system, in line with existing EU product
safety legislation. Classification depends on the func-
tion the AI system performs and on the specific pur-
pose and modalities for which the system is used.
“The legal uncertainty surrounding [regulation] and
the lack of budget to hire specialized lawyers or mul-
tidisciplinary teams still are significant barriers for a
flourishing AI startup and scale-up ecosystem,” Kop
says. “The question remains whether the AI Act will im-
prove or worsen the startup climate in the EU.”
The EC will determine which AI gets classified as
“high risk” using criteria that are still under debate,
creating a list of examples of high-risk systems to help
guide judgment.
“It will be a dynamic list that contains various types of
AI applications used in certain high-risk industries, which
means the rules get stricter for riskier AI in healthcare and
defense than they are for AI apps in tourism,” Kop says.
“For instance, medical AI is [classified as] high risk to pre-
vent direct harm to patients due to AI errors.”
He notes there is still controversy about the criteria and
definition of AI that the draft uses. Some commentators
argue it should be more technology-specific, aimed at
certain riskier types of machine learning, such as deep
unsupervised learning or deep reinforcement learning.
“Others focus more on the intent of the system, such as
social credit scoring, instead of potential harmful results,
such as neuro-influencing,” Kop adds. “A more detailed
classification of what ‘risk’ entails would thus be welcome
in the final version of the act.”
Facial Recognition as a High-Risk
Technology
Joseph Carson, chief security scientist and advisory CISO
at Delinea, participated in several of the talks around the
act, including as a subject matter expert in the use of AI
in law enforcement and articulating the concerns around
security and privacy.
September 2022 14
Previous Next
EU DEBATES AI ACT TO PROTECT HUMAN RIGHTS, DEFINE HIGH-RISK USES
The EU AI Act, he says, will mainly affect those organi-
zations that already collect and process personal identi-
fiable information. Therefore, it will impact how they use
advanced algorithms in processing the data.
“It is important to understand the risks if no regulation
or act is in place and what the possible impact [is] if or-
ganizations abuse the combination of sensitive data and
algorithms,” Carson says. “The future of the Internet is a
scary place, and the enforcement of the EU AI Act allows
us to embrace the future of the Internet using AI with both
responsibility and accountability.”
Regarding facial recognition, he says the technology
should be regulated and controlled.
“It has many amazing uses in society, but it must be
something you opt in and agree to use; citizens must
have a choice,” he says. “If no act is in place, we will
see a significant increase in deepfakes that will spiral
out of control.”
Malin Strandell-Jansson, senior knowledge expert at
McKinsey & Co, says facial recognition is one of the most
debated issues in the draft act, and the final outcome is
not yet clear.
In its draft format, the AI Act strictly prohibits the use
of real-time remote biometric identification in publicly
accessible spaces for law enforcement purposes, as it
poses particular risks for fundamental rights — notably
human dignity, respect for private and family life, protec-
September 2022 15
tion of personal data, and nondiscrimination.
Strandell-Jansson points out a few exceptions, includ-
ing use for law enforcement purposes for the targeted
search for specific potential victims of crime, including
missing children; the response to the imminent threat of a
terror attack; or the detection and identification of perpe-
trators of serious crimes.
“Regarding private businesses, the AI Act considers all
emotion recognition and biometric categorization sys-
tems to be high-risk applications if they fall under the use
cases identified as such — for example, in the areas of
employment, education, law enforcement, migration, and
border control,” she explains.
As such, potential providers would have to subject such
AI systems to transparency and conformity obligations
before putting them on the market in Europe.
The Time to Act on AI Is Now
Dr. Sohrob Kazerounian, AI research lead at Vectra, an AI
cybersecurity company, says the need to create a regula-
tory framework for AI has never been more pressing.
“AI systems are rapidly being integrated into products
and services across wide-ranging markets,” he says. “Yet
the trustworthiness and interpretability of these systems
can be rather opaque, with poorly understood risks to
users and society more broadly.”
While some existing legal frameworks and consumer
Previous Next
EU DEBATES AI ACT TO PROTECT HUMAN RIGHTS, DEFINE HIGH-RISK USES
protections may be relevant, applications that use AI are
sufficiently different enough from traditional consumer
products that they necessitate fundamentally new legal
mechanisms, he adds.
The overarching goal of the bill is to anticipate and miti-
gate the most critical risks resulting from the use and failure
of AI, with actions ranging from banning systems deemed
to have “unacceptable risk” altogether to heavy regulation
of “high-risk” systems. Another, albeit less-noted, conse-
quence of the framework is that it could provide clarity and
certainty to markets about what regulations will exist and
how they will be applied.
“As such, the regulatory framework may in fact result in
increased investment and market participation in the AI
sector,” Kazerounian says.
Limits for Deepfakes and Biometric
Recognition
By addressing specific AI use cases, such as deepfakes
and biometric or emotional recognition, the AI Act is hop-
ing to ameliorate the heightened risks such technologies
pose, such as violation of privacy, indiscriminate or mass
surveillance, profiling and scoring of citizens, and manip-
ulation, Strandell-Jansson says.
“Biometrics for categorization and emotion recognition
have the potential to lead to infringements of people’s
privacy and their right to the protection of personal data
as well as to their manipulation,” she says. “In addition,
there are serious doubts as to the scientific nature and
reliability of such systems.”
The bill would require people to be notified when they
encounter deepfakes, biometric recognition systems, or
AI applications that claim to be able to read their emo-
tions. Although that’s a promising step, it raises a couple
of potential issues.
Overall, Kazerounian says it is “undoubtedly” a good
start to require increased visibility for consumers when
they are being classified by biometric data and when they
are interacting with AI-generated content rather than real
humans or real content.
“Unfortunately, the AI act specifies a set of application
areas within which the use of AI would be considered
high-risk, without necessarily discussing the risk-based
criteria that could be used to determine the status of fu-
ture applications of AI,” he says. “As such, the seemingly
ad-hoc decisions about which application areas are con-
sidered high-risk simultaneously appear to be too specif-
ic and too vague.”
Current high-risk areas include certain types of biomet-
ric identification, operation of critical infrastructure, em-
ployment decisions, and some law enforcement activi-
ties, he explains.
“Yet it isn’t clear why only these areas were considered
high-risk and furthermore doesn’t delineate which appli-
September 2022 16
cations of statistical models and machine learning sys-
tems within these areas should receive heavy regulatory
oversight,” he adds.
Possible Groundwork for Similar US Law
It is unclear what this act could mean for a similar law in
the US, Kazerounian says, noting that it has now been
more than half a decade since the passing of GDPR, the
EU law on data regulation, without any similar federal
laws following in the US — yet.
“Nevertheless, GDPR has undoubtedly influenced the
behavior of multinational corporations, which have ei-
ther had to fracture their policies around data protec-
tions for EU and non-EU environments or simply have a
single policy based on GDPR applied globally,” he says.
“In any case, if the US decides to propose legislation on
the regulation of AI, at a minimum it will be influenced by
the EU act.”
About the Author: Nathan Eddy is a freelance journalist and
award-winning documentary filmmaker specializing in IT security,
autonomous vehicle technology, customer experience technology,
and architecture and urban planning. A graduate of Northwestern
University’s Medill School of Journalism, Nathan currently lives in
Berlin, Germany.
Previous Next
SPONSORED CONTENT

DEEP INSTINCT PERSPECTIVES

Harnessing the Power of Deep Learning

to Speed Up Threat Detection
AI is the future of security — but not all AI is created equal.
By Chuck Everette, Director of Cybersecurity Advocacy, Deep Instinct

T
he pace of the cybersecurity arms race is accelerating. Attacks are increasing
exponentially, and breaches are on the rise. As soon as a defensive measure
becomes mainstream, the threat actors develop a new workaround.
It’s becoming clear that artificial intelligence (AI) is the future of cybersecurity tools.
But to stay ahead of the attack groups, AI-based solutions need to rapidly evolve and
be positioned to predict and prevent attacks rather than react to them. A form of AI
known as deep learning (DL) has emerged as the most effective way of delivering a
prevention-first approach to take back the advantage from attackers.
So, what is deep learning, and how does it differ from other forms of AI?

AI Defined
The term AI has become an overused sales buzzword, muddying the waters for enter-
prises seeking the best solution for their needs. While often applied indiscriminately, AI
is an umbrella term that can be split into three distinct types: basic AI, reactive machine
learning (ML), and proactive DL.

September 2022 17
Previous Next
HARNESSING THE POWER OF DEEP LEARNING TO SPEED UP THREAT DETECTION
In its simplest form, basic AI involves human-created al-
gorithms relying on the use of signatures and scripts. This
form requires constant human intervention to stay up to
date and can be applied only to recognizing and flagging
known threat patterns and behaviors. Basic AI is com-
monly deployed in tools such as antivirus software, and
it’s prone to a high false-positive rate and critical delays
due to its need to check threat signatures against a cloud
database.
The Strengths and Weaknesses of Reactive
Machine Learning
ML takes a step beyond basic AI by enabling greater be-
havioral analysis, pattern recognition, and event correla-
tion, and when applied to cybersecurity solutions it most
often results in a reactive security posture.
With ML, the algorithms are trained on only 1% to 2%
of all available data, which must be continually updated
with manually engineered datasets that enable them to
examine behaviors on the endpoint after malware execu-
tion. That said, the fully human-led training process used
for standard machine learning requires a high degree of
expertise, can be quite time intensive, and potentially can
introduce human bias.
ML-powered tools have come to be particularly useful in
reactive, post-execution threat detection and analytics as
they can do the heavy lifting for security operations center
teams and comb through large volumes of data far faster
than any human.
However, the technology still has some serious short-
comings. Manual training can only use currently available
data from known attacks and behaviors. Machine learning
requires samples of existing threats so that humans can
manually label and train the algorithm to detect and re-
spond to threats. This is ML’s Achilles’ heel. The vast major-
ity of newer threats are missed because ML has not trained
on them in the past — an extremely dangerous blind spot.
Furthermore, false-positive rates are usually around 1% to
2% — and this becomes a huge problem when there are
tens of thousands of alerts a day.
But, most crucially, ML is a firmly reactive application of
AI. Threats need to have actively executed within the net-
work environment before they can be detected. Consider-
ing that the fastest ransomware can begin spreading and
damaging the network in a matter of seconds, this reactive
stance is increasingly not fast enough to stop most attacks
before the attackers get inside an environment and start to
move laterally, thereby creating backdoors, stealing data,
and causing other damage.
The Deep Learning Difference
As the most advanced form of AI, deep learning (DL) is
extremely fast at detecting and preventing these threats.
DL takes a proactive approach to security, focusing
September 2022 18
Previous Next
HARNESSING THE POWER OF DEEP LEARNING TO SPEED UP THREAT DETECTION

on identifying incoming threats the moment they hit the ingly critical to getting out ahead of rapidly innovating
environment and blocking them before they can execute. threat actors. The speed, power, and accuracy of deep
This proactive approach is facilitated by a more learning provides the critical edge that enterprises need
sophisticated application of AI. Rather than being trained to regain the advantage.
on data that already has been manually labeled as benign
or malicious, DL solutions are fed 100% raw data, and they About Deep Instinct
will identify the correct patterns independently. Deep Instinct takes a prevention-first approach to stopping ran-
This results in a large, multilayered model that utilizes all somware and other malware using a deep learning cyberse-
available data rather than a human-determined subset. A curity framework. We predict and prevent known, unknown,
properly fully trained DL solution is, therefore, less affected and zero-day threats in <20 milliseconds, 750X faster than
by bias, and it gains far greater pattern recognition than a the fastest ransomware can encrypt. Deep Instinct has >99%
standard ML solution. The more organic training process zero-day accuracy and promises a <0.1% false-positive rate.
results in a greater degree of accuracy and renders deep The Deep Instinct Prevention Platform provides complete,
learning more likely to identify attack techniques and multilayered protection against threats across hybrid environments.
tactics missed by ML models. For more, visit http://www.deepinstinct.com.
Crucially, this means a much greater ability to identify
zero-day threats. Even a DL solution that has not been
updated for the better part of a year can identify and block
a previously unknown threat.
Alongside this, DL-based cybersecurity solutions can op-
erate at a speed and with accuracy that can’t be touched
by a ML-based solution. The best DL-powered solutions
can identify threats in less than 20 milliseconds, far faster
than even the swiftest known ransomware can execute.
The false-positive rate can also be as little as 0.1%.
As the pace of the cyber arms race continues to accel-
erate, proactive, prevention-first security will be increas-

September 2022 19
Previous