Explore the intersection of big data security and artificial intelligence in this 39-minute conference talk by Dr. Murat Kantarcioglu from the University of Texas at Dallas. Delve into the changing landscape of the big data revolution and its associated challenges. Examine hardware-supported approaches for efficient oblivious data processing and their applications in basic data science. Compare experimental evaluations with ObliVM and investigate the trade-offs between privacy and robustness in federated learning. Analyze backdoor attacks in the context of federated learning, including experiments and comparisons with other defenses. Conclude with insights on federated learning poisoning attacks and explore techniques for attacking models to enhance privacy and fairness, such as preventing gender prediction in image classifiers.
Overview
Syllabus
Intro
Big Data Revolution: Changing Landscape
Challenges for Big Data
Other Approach: Use Hardware Support for Efficient Oblivious Data Processing
How to Support Data Obliviousness ??
Support for Basic Data Science
Experimental Evaluation
Comparison with ObliVM
Federated Learning: Privacy vs Robustness
Backdoor Attacks in FL context
Overview
Experiments
Comparison with Other Defenses - IID
Conclusion: FL Poisoning Attacks
Attacking models to improve privacy and fairness
Example: Attacking Image Classifiers
Domain constraint Example
Example: Prevent Gender Prediction
Change Images Using Glasses
Questions?
Taught by
CAE in Cybersecurity Community
