NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips

Valerio Venceslai, Alberto Marchisio, Ihsen Alouani, Maurizio Martina, Muhammad Shafique

Abstract

Due to their proven efficiency, machine-learning systems are deployed in a wide range of complex real-life problems. More specifically, Spiking Neural Networks (SNNs) emerged as a promising solution to the accuracy, resource-utilization, and energy-efficiency challenges in machine-learning systems. While these systems are going mainstream, they have inherent security and reliability issues. In this paper, we propose NeuroAttack, a cross-layer attack that threatens the SNNs integrity by exploiting low-level reliability issues through a high-level attack. Particularly, we trigger a fault-injection based sneaky hardware backdoor through a carefully crafted adversarial input noise. Our results on Deep Neural Networks (DNNs) and SNNs show a serious integrity threat to state-of-the art machine-learning techniques.

Introduction

As Spiking Neural Networks (SNNs) transition from academic research to mainstream deployment in energy-efficient hardware, such as Intel’s Loihi, their security profile becomes a critical concern. While traditional Deep Neural Networks (DNNs) are known to be vulnerable to adversarial examples and fault-injection attacks, the security landscape for SNNs is less understood.

NeuroAttack is a novel cross-layer attack that bridges the gap between software-level adversarial noise and hardware-level fault injection. It introduces a threat where a hardware vulnerability—a Trojan—is embedded in the circuit during manufacturing but remains dormant until activated by a specific high-level input pattern. This allows a remote attacker to compromise a system's integrity without needing physical access to the device during operation.

The NeuroAttack Framework

The NeuroAttack framework consists of three main stages designed to exploit the specific characteristics of SNNs:

  1. Vulnerability Analysis: Using a gradient-based search algorithm, researchers identify the most "sensitive" bits in the network's weights. While SNNs are resilient to random bit-flips, they are incredibly fragile when specific bits are targeted.
  2. Hardware Trojan Design: A lightweight hardware Trojan is designed using a 2-way multiplexer and an inverter. When triggered, it flips the target weight bit. To adapt this for SNNs, which communicate via timing and frequency, a counter tracks spike rates to determine when to activate the flip.
  3. Trigger Generation: An optimization loop crafts a specific noise pattern (the trigger) that maximizes the activation of a single "target neuron." When this neuron's output exceeds a threshold, it activates the hardware Trojan.
The NeuroAttack mechanism: A target neuron's activity is monitored by a counter and comparator to trigger a hardware Trojan that flips weight bits.

Experimental Evaluation and Key Results

The authors evaluated NeuroAttack using standard computer vision benchmarks, including MNIST (for MLP and LeNet-5) and CIFAR-10 (for complex CNNs). The results highlight a stark contrast between surgical precision and brute-force attacks. While SNNs can withstand thousands of random errors, flipping just 30 targeted bits can reduce accuracy to near zero. On the more complex CIFAR-10 dataset, only 4 surgical bit-flips were needed to crash the model.

Impact of targeted bit-flips on model accuracy: Only a few surgical bit-flips are required to significantly degrade performance across different architectures and datasets.

Key findings include:

  • Effectiveness: On MNIST, the attack achieved a zero false-activation rate on clean images while successfully triggering on 100% of malicious inputs.
  • Minimal Overhead: The hardware cost is approximately 540 transistors, an incredibly small footprint in modern chips with billions of transistors, making detection difficult.
  • Cross-Domain Transfer: Vulnerabilities identified in DNNs could be effectively transferred to SNNs, proving that the conversion to discrete spikes does not inherently provide security against these attacks.

Discussion and Future Work

The core innovation of NeuroAttack is using an internal neuron as a "remote control" for a hardware Trojan. By optimizing an input pattern to over-activate one specific neuron, the attacker creates a trigger that bypasses traditional hardware security checks. The attack prioritizes stealth through minimal hardware overhead and mathematically optimized, nearly invisible noise triggers.

However, the attack currently assumes "grey-box" access to the architecture during the design or manufacturing phase. Furthermore, the trigger is specific to the network's weights; if the weights are updated via on-chip learning, the original trigger may become ineffective. Future research should focus on hardware-software co-design for security, potentially developing monitors to detect unusual internal neuron activity that does not correlate with typical input features.

Summary updated

Abstract

Due to their proven efficiency, machine-learning systems are deployed in a wide range of complex real-life problems. More specifically, Spiking Neural Networks (SNNs) emerged as a promising solution to the accuracy, resource-utilization, and energy-efficiency challenges in machine-learning systems. While these systems are going mainstream, they have inherent security and reliability issues. In this paper, we propose NeuroAttack, a cross-layer attack that threatens the SNNs integrity by exploiting low-level reliability issues through a high-level attack. Particularly, we trigger a fault-injection based sneaky hardware backdoor through a carefully crafted adversarial input noise. Our results on Deep Neural Networks (DNNs) and SNNs show a serious integrity threat to state-of-the art machine-learning techniques.
BibTeX
@inproceedings{venceslai2020neuroattack,
  author = {Valerio Venceslai and Alberto Marchisio and Ihsen Alouani and Maurizio Martina and Muhammad Shafique},
  title = {{NeuroAttack: Undermining Spiking Neural Networks Security through Externally Triggered Bit-Flips}},
  booktitle = {IJCNN},
  year = {2020},
  url = {https://openreview.net/forum?id=FiVc0ke7HW_},
}
Figure