TrISec: Training Data-Unaware Imperceptible Security Attacks on Deep Neural Networks

Faiq Khalid, Muhammad Abdullah Hanif, Semeen Rehman, Rehan Ahmed, Muhammad Shafique

Introduction

As autonomous vehicles (AVs) transition from research labs to public roads, their reliance on Deep Neural Networks (DNNs) for critical tasks like traffic sign recognition introduces significant security risks. Adversarial attacks, which involve subtly altering input data to cause misclassification, pose a direct threat to passenger safety and system reliability.

What Problem Does This Paper Address?

Existing adversarial attacks often suffer from two primary limitations that hinder their real-world effectiveness. First, many "data poisoning" techniques require access to the original training dataset—a "white-box" requirement that is often unrealistic for external attackers. Second, these attacks frequently introduce visible noise or structural anomalies that can be detected by human observers or automated security systems using statistical checks like Correlation Coefficients (CR) or the Structural Similarity Index (SSI).

Overview of security threats in the machine learning pipeline for autonomous vehicles, contrasting training-time and inference-time attacks.

The authors of TrISec aimed to overcome these hurdles by developing an attack framework that is entirely training data-unaware and generates perturbations so subtle they remain "invisible" to both human and machine scrutiny, bypassing multi-level security filters unnoticed.

The TrISec Methodology

TrISec employs an iterative optimization process leveraging backpropagation on pre-trained DNNs. The framework is designed to satisfy two competing objectives:

  1. Targeted Misclassification: The algorithm identifies a target "fake" class (e.g., misidentifying a "Stop" sign as a "60km/h" speed limit). It then calculates the precise pixel perturbations ($\delta$) required to minimize the model's cost function for that target label, forcing a high-confidence error.
  2. Guaranteed Imperceptibility: To ensure the attack remains undetected, TrISec optimizes for two mathematical similarity metrics. It maximizes the Correlation Coefficient (CR) to maintain the global relationship between the original and altered images, and it maximizes the Structural Similarity Index (SSI) to prevent localized noise clusters that would be visible to the eye.

By iteratively updating the perturbation, TrISec finds the "sweet spot" where the image successfully tricks the model while maintaining near-perfect similarity to the original reference.

Key Results and Insights

The researchers evaluated TrISec using the German Traffic Sign Recognition Benchmark (GTSRB) and the VGGNet architecture. The results demonstrate a significant breakthrough in adversarial robustness testing:

  • High-Confidence Deception: TrISec successfully caused a "Stop" sign to be misclassified as a "60km/h" sign with approximately 95% confidence.
  • Near-Perfect Similarity: The generated attack images achieved a CR of ~1 and an SSI of 0.999, effectively bypassing standard security filters.
  • Breaking the Trade-off: Traditional methods like FGSM and L-BFGS often fail to cause misclassification when constrained by high similarity requirements (CR > 0.95 and SSI > 0.99). TrISec, however, maintains high success rates even under these strict bounds.
Scenario III: An imperceptible noise attack where a 'Stop' sign is misclassified as a '60km/h' sign despite appearing unchanged to the human eye.

A core insight of the paper is that the pre-trained model's weights themselves contain sufficient information to generate effective attacks, rendering access to training data unnecessary. Furthermore, the study highlights that a single similarity metric is insufficient for security; multi-metric optimization is required to ensure true imperceptibility.

Limitations and Future Work

While highly effective, TrISec is computationally intensive, often requiring tens of thousands of iterations to converge. Additionally, the current framework assumes a "white-box" scenario where the attacker has access to the model's parameters. Future research will likely explore "black-box" applications and the development of more robust, security-aware DNN architectures capable of defending against such sophisticated, subtle perturbations.

Summary updated

Abstract

Most of the data manipulation attacks on deep neural networks (DNNs) during the training stage introduce a perceptible noise that can be catered by preprocessing during inference, or can be identified during the validation phase. There-fore, data poisoning attacks during inference (e.g., adversarial attacks) are becoming more popular. However, many of them do not consider the imperceptibility factor in their optimization algorithms, and can be detected by correlation and structural similarity analysis, or noticeable (e.g., by humans) in multi-level security system. Moreover, majority of the inference attack rely on some knowledge about the training dataset. In this paper, we propose a novel methodology which automatically generates imperceptible attack images by using the back-propagation algorithm on pre-trained DNNs, without requiring any information about the training dataset (i.e., completely training data-unaware). We present a case study on traffic sign detection using the VGGNet trained on the German Traffic Sign Recognition Benchmarks dataset in an autonomous driving use case. Our results demonstrate that the generated attack images successfully perform misclassification while remaining imperceptible in both “subjective” and “objective” quality tests.
BibTeX
@inproceedings{khalid2019trisec,
  author = {Faiq Khalid and Muhammad Abdullah Hanif and Semeen Rehman and Rehan Ahmed and Muhammad Shafique},
  title = {{TrISec: Training Data-Unaware Imperceptible Security Attacks on Deep Neural Networks}},
  booktitle = {IOLTS},
  year = {2019},
  doi = {10.1109/iolts.2019.8854425},
}
Figure