Hardware Trojan Attacks on Neural Networks

### The Threat of Hardware Trojans in Deep Learning Hardware At 00:00, Joseph Clements presents research on the vulnerabilities of dedicated neural network processing chips at the AI Village. Clements examines how malicious modifications introduced during silicon design or manufacturing compromise deep learning models running on edge devices.

### Embedded Systems Constraints and Specialized Accelerators At 04:39, Clements reviews the architectural trade-offs of deploying deep neural networks on battery-constrained edge platforms. To achieve real-time performance, mobile systems rely heavily on dedicated ASIC and FPGA hardware accelerators whose complex silicon layouts are difficult to verify.

### Supply Chain Risks in Semiconductor Fabrication At 09:49 and 14:50, the talk explores semiconductor supply chain vectors. Because modern chip design relies on outsourced third-party fabrication foundries and commercial intellectual property (IP) cores, adversaries can insert dormant hardware trojans into mask sets prior to fabrication without detection.

### Trigger Mechanics and Activation Manipulation in Silicon At 19:34, Clements details how hardware trojans operate in silicon. By altering arithmetic logic units or modifying internal activation function circuits upon receiving subtle hardware or physical triggers, the trojan forces targeted model misclassifications while behaving normally under diagnostic test suites.

(AI summary — Gemini Flash)