BARE-AI: Bit-Flip Attack Resilience in AI Hardware through Built-in Performance Monitors
BARE-AI detects and mitigates bit-flip attacks on neural accelerators using lightweight on-chip monitors with under 4% area overhead.
BARE-AI is a runtime framework that detects, localizes, and mitigates memory bit-flip attacks on DNN accelerators during inference. Lightweight AI Performance Counters capture per-layer activation statistics that an on-chip ensemble detector (PULSE) scores; an Activation Shift Index localizes faults and a z-score repair resets anomalous weights toward clean layer statistics. Across CNNs, Vision Transformers, and LLMs under random, targeted, adaptive, and magnitude-based flips, detection reaches up to 98% on vision models and 74% to 95% on language models, with near-clean recovery for CNNs and ViTs and partial recovery for LLMs. Synthesized at 28nm, monitoring costs under 3% energy, under 4% area, and about 10% latency, configurable to about 6%, independent of how many flips are tolerated.
- AI Performance Counters track per-layer sparsity, entropy, kurtosis, and spectral shift.
- Detection reaches up to 98% on vision models and 74–95% on language models.
- Z-score repair restores near-clean accuracy for CNNs and ViTs; LLMs recover only partly.
- At 28nm, overhead is under 3% energy, under 4% area, and about 6–10% latency.
- Products
- BARE-AI
From the original article
1 min read at arxiv.orgDeep Neural Networks (DNNs) are integral to many safety critical systems, yet they remain highly vulnerable to bit-flip attacks (BFAs), where a few memory level perturbations can drastically degrade accuracy. Existing defenses incur significant hardware overhead, depend on retraining, or fail against targeted flips. We propose BARE-AI, a runtime framework that detects, localizes, and mitigates BFAs during inference.
By Habibur Rahaman · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.