Source-linked AI summary
Hardware Trojan Threats to Multi-Chiplet Photonic Neural Network Accelerators
Sudeep Pasricha
TL;DR
MCPNAs offer scalable, energy-efficient AI acceleration but introduce Hardware Trojan attack surfaces through distributed architecture and third-party chiplets. This paper examines confidentiality, integrity, and availability threats and discusses cross-layer countermeasures. It highlights optical snooping, microring and thermal computation attacks, and coordinated denial-of-service threats, alongside authentication, resilient training, monitoring, and trust-aware routing defenses.
Problem
MCPNAs’ distributed architecture and reliance on third-party chiplets create Hardware Trojan risks across model confidentiality, computation integrity, and system availability.
Method
The paper analyzes MCPNA Hardware Trojan threats across confidentiality, integrity, and availability, then discusses photonic, algorithmic, runtime, and routing countermeasures.
Results
The analysis identifies optical-interconnect snooping, microring actuation and thermal attacks, and coordinated denial-of-service attacks as major MCPNA threat classes.
Takeaways & Limitations
Trustworthy MCPNA operation requires holistic cross-layer security combining photonic authentication, resilient model training, runtime monitoring, and trust-aware routing.
Abstract
from arXiv · showhide
Multi-chiplet photonic neural network accelerators (MCPNAs) combine photonic computation, photonic communica-tion, and heterogeneous chiplet integration to enable scalable and energy-efficient AI acceleration. However, their distributed archi-tecture and reliance on third-party chiplets introduce significant hardware security risks. This paper examines Hardware Trojan (HT) threats to MCPNAs across three dimensions: confidentiality, integrity, and availability.
I. INTRODUCTION
MCPNAs scale photonic neural acceleration through heterogeneous chiplet integration and photonic interconnects, but distributed components and third-party supply chains expand Hardware Trojan exposure.
- Architecture: 2.5D chiplet integration scales photonic neural accelerators beyond monolithic designs by linking compute and memory chiplets through a silicon-photonic interposer.Wavelength-division multiplexing connects distributed photonic components, including microring resonators, gateways, routing structures, photodetectors, and tuning circuits.
- Threat model: HTs may enter compute, memory, gateway, or interposer chiplets during design, fabrication, or integration.The threat model includes independent or colluding attacks across the MCPNA supply chain.
- Security objectives: Photonic sensitivity to wavelength, temperature, and process variations makes Trojan-induced perturbations difficult to detect.The security objectives are protecting model confidentiality, computation integrity, and system availability.
II. CONFIDENTIALITY ATTACKS
MCPNA optical interconnects expose model data to covert Hardware Trojan snooping, while photonic process-variation signatures offer a basis for authentication and encryption defenses.
- Attack mechanism: HTs in microring tuning circuits can retune detector resonators toward neighboring wavelengths and covertly snoop shared WDM channels.The attack can operate with little impact on legitimate communication.
- Exposed information: Compromised optical channels can reveal memory-fetched weights, inter-chiplet activations, and proprietary model parameters.Such leakage threatens model confidentiality and intellectual-property protection as third-party chiplet use increases.
- Countermeasures: Process-variation signatures can derive unclonable authentication and encryption keys that protect photonic communications from snooping.The same fingerprints could support chiplet authentication, trusted onboarding, and secure communication protocols.
III. INTEGRITY ATTACKS
Integrity attacks corrupt photonic neural computations through microring actuation and thermal effects, with distributed layers allowing errors to propagate across the inference pipeline.
- Actuation attacks: HTs in microring actuation circuitry can force off-resonance states that corrupt weight or activation encoding and produce incorrect computations.Distributed neural-network layers allow these errors to propagate through later computations and inter-chiplet communications.
- Thermal attacks: Thermal hotspot attacks shift multiple microring resonance wavelengths through localized heating and can affect neighboring devices or chiplets through thermal coupling.Dense photonic MAC arrays sharing an interposer substrate extend the attack beyond the compromised device.
- Impact: Even a small fraction of compromised microring resonators can cause significant accuracy degradation, with thermal attacks often exceeding actuation attacks.Thermal hotspots corrupt parameter clusters rather than isolated weights.
- Countermeasures: L2 regularization and noise-aware training improve tolerance to Hardware Trojan perturbations.Proposed future defenses combine resilient training with runtime thermal monitoring, resonance tracking, redundancy, and chiplet-level fault containment.
IV. AVAILABILITY ATTACKS
MCPNA communication resources are vulnerable to coordinated Hardware Trojan denial-of-service attacks, while trust-aware and adaptive routing can reduce exploitation of critical optical paths.
- Attack mechanisms: Malicious chiplets can coordinate distributed denial-of-service attacks by injecting traffic, monopolizing wavelength channels, or overwhelming photonic gateway interfaces.These attacks target high-bandwidth inter-chiplet communication resources and critical optical paths.
- System impact: Predictable memory-to-compute routes enable congestion hotspots that increase latency and reduce accelerator utilization.Manipulating gateway or wavelength allocation in reconfigurable interposers can further reduce effective bandwidth.
- Countermeasures: Trust-aware routing uses chiplet trust levels, path obfuscation, randomized route selection, and load balancing to reduce exploitation of critical paths.Photonic extensions include dynamic wavelength hopping, adaptive interposer reconfiguration, wavelength assignment, and gateway selection.
V. CONCLUSIONS
MCPNAs face Hardware Trojan risks across confidentiality, integrity, and availability, motivating holistic cross-layer defenses for trustworthy photonic chiplet AI systems.
- HTs can compromise confidentiality through photonic interconnect snooping, integrity through microring actuation and thermal attacks, and availability through coordinated denial-of-service attacks.
- Emerging countermeasures include photonic authentication, resilient model training, runtime monitoring, and trust-aware routing.
- Future protection requires cross-layer defenses spanning trusted chiplet onboarding, secure communication, resilient training, anomaly detection, thermal management, and network orchestration.
- These defenses are increasingly important as photonic AI platforms scale to LLMs, graph learning, and vision-based generative AI workloads.