RandEye: On-Sensor Stochastic Image Transformation for Backdoor-Resistant Edge Inference

The rising threat of backdoor attacks in artificial intelligence (AI) models poses significant risks across diverse applications, including medical diagnostics, autonomous navigation, and surveillance systems. Existing software-based defenses, though effective, are computationally demanding and impractical for low-latency, low-power edge AI inference. This work presents RandEye, the first hardware-based backdoor defense directly embedded into a CMOS image sensor (CIS) and adaptable to various CIS designs. RandEye applies stochastic transformations to captured images directly on the CIS to disrupt malicious input patterns crafted by adversaries, thus effectively deactivating backdoor triggers in compromised models. We introduce the design techniques of hardware-efficient affine transformations and pixel reverse transformation to achieve a lightweight on-sensor defense. RandEye integrated on a 256×256 pixel array is evaluated to achieve a low power profile of 0.35 mW and requires only 1.8% area overhead, while supporting frame rates exceeding 30 FPS for downstream image classification. For backdoored models, RandEye reduces the attack success rate to below 5% while maintaining model accuracy (ACC). When paired with model fine-tuning, RandEye further improves ACC by around 8.8% while still robustly suppressing ASR.