Presentation
Classification Prediction Attack over Sparsity-aware Embedded Neural Networks Using Timing-based Side-Channel Leakage
DescriptionThis work analyzes the side-channel security vulnerabilities associated with sparsity-aware-based optimizations for Neural Network platforms. In particular, optimizations like gating the execution of sparse multiplications give rise to timing side-channel attacks. We propose a classification prediction attack that uses the timing side-channel information to mimic the prediction result of the underlying Neural Network implementation. We demonstrated our techniques for the CIFAR-10 classification task over timing models for Neural network accelerators and microcontroller platforms across different dataflow algorithms and processing loads. Our evaluation results show that the classification results are predictable with significantly high accuracy.
Event Type
Work-in-Progress Poster
TimeWednesday, July 12th6:00pm - 7:00pm PDT
LocationLevel 2 Lobby
AI
Autonomous Systems
Cloud
Design
EDA
Embedded Systems
RISC-V
Security