Professional Documents
Culture Documents
IIT KHARAGPUR
Every organization wants to protect its valuable digital assets like data and information, but
what about AI algorithms? Successful exploitation of AI algorithms can cause financial loss,
reputational damage, loss of competitive advantage, and loss of intellectual property. For
example, a malicious breach of pricing algorithm for an e-commerce company can result in
attackers causing financial loss to the retailers/sellers and negatively impacting the
platform's brand reputation.The majority of organizations are struggling to address the
challenge of securing models. The first aspect of protecting the asset is understanding if
they can be attacked. In this case, extracting the video-based DL algorithms.
| HiGH PREP
PROBLEM STATEMENT
Develop an efficient common strategy and relevant implementation to extract the
video-based models in the black box and grey box setting across the following 2
problem statements.
1.Action Classification
Model Extraction for Swin-T Model for Action Classification on
Kinetics-400 dataset.
Download the model from here-
https://github.com/SwinTransformer/Video-Swin-Transformer
2.Video Classification
Model Extraction for MoViNet-A2-Base Model for Video Classification on Kinetics-
600 dataset
Download the model from here-
https://tfhub.dev/tensorflow/movinet/a2/base/kinetics- 600/classification/3
Blackbox Setting
Do not use any relevant data set available and use synthetic or generated data
without using the Kinetics series dataset. Also, do not use the same model
architecture as the original model to train the extracted model.
Greybox Setting
You can use 5% of original data (balanced representation of classes) as a starting
point to generate the attack dataset. Also, do not use the same model architecture as
the original model to train the extracted model.
| HiGH PREP
GUIDELINES
The competition will use a private leaderboard, and no results will be announced
intermittently. (detailed mechanism will be contributed)
All teams will submit a score in the specified format and its sub-component (Px
black/grey accuracy, no. of queries) at the end of week 3 and week 4.
After week 4, each team will also be asked to give a small presentation explaining
their work.
SUBMISSIONS
The evaluation will be conducted in three steps-
Please note that the Final presentation should wrap up within 15 minutes which will be
followed by Q&A (5 min).
| HiGH PREP
EVALUATION CRITERIA
The results of extraction will be measured using the following evaluation metric
| HiGH PREP