ECE PhD Dissertation Defense by: Dev Mehta

Monday, November 23, 2026
11:00 a.m. to 1:00 p.m.
Floor/Room #
AK 218
Dev Mehta

Title:

Physical Security in AI Accelerators

 

Abstract:

With the increasing deployment of artificial intelligence (AI) on physical devices, the security of its intellectual property (IP) has become increasingly important. Although the security paradigms employed to protect these circuits have been shown to be robust in simulation, they frequently exhibit weaknesses once they are implemented in hardware. These weaknesses are rooted in the inherent physics of CMOS-based circuits, which renders it considerably more difficult for designers to create circuits that withstand a wide variety of attack vectors and to avoid design choices that inadvertently weaken their designs.

This dissertation investigates a range of side-channel analysis techniques in order to expose the vulnerabilities of such designs, and it makes two primary contributions. First, we demonstrate physical attacks that break several protection schemes that can be employed to protect neural networks (NNs). Second, we study the impact of physical design choices on security from both the attack and the protection perspectives, with particular emphasis on timing analysis. FPGA implementations, in combination with pre-silicon analysis, reveal the influence of timing on glitches and show how an elevated operating temperature can render a masked NN implementation vulnerable to a first-order attack. Beyond the power side channel, we further demonstrate that optical side channels can be exploited to attack AI accelerators at different levels of protection. Overall, our results show that the security guarantees of masked and cryptographically protected AI hardware, which hold in abstract security models, such as the probing model for masking and the cryptographic security models of garbled and universal circuits, do not carry over to the physical implementation, since timing differences below the clock-cycle abstraction and backside optical access expose sensitive values to an adversary with physical access.

 

Research Advisor:

Prof. Fatemeh (Saba) Ganji

ECE Department, WPI

 

Committee Members:

Prof. Patrick Schaumont

ECE Department, WPI

Prof. Domenic Forte

ECE Department, University of Florida