TrustNoC: Trustworthy On-Interposer NoC for AI Backdoor Mitigation in Chiplet Systems

Abstract

The rise of chiplet-based AI accelerators creates a critical security gap: untrusted, black-box components can introduce malicious backdoors, yet their nature renders traditional software defenses impractical. To close this gap, we propose TrustNoC, the first trustworthy interposer architecture that embeds a novel backdoor filter ensemble (BFE) into its network-onchip routers to filter malicious data in-transit without modifying the AI models or the chiplets themselves. Our evaluation shows TrustNoC effectively suppresses backdoor attacks, reducing the attack success rate to 0.5% while maintaining model accuracy above $90 \%$. This robust hardware-level security is achieved with a low footprint, requiring <10% of the total interposer area to maintain high yield. Thus, TrustNoC establishes the silicon interposer as a new hardware defense layer, offering a practical solution for secure AI inference in heterogeneous chiplet systems.