Challenging the adversarial robustness of DNNs based on error-correcting output codes

Bowen Zhang, Benedetta Tondi, Xixiang Lv, Mauro Barni

Abstract

The existence of adversarial examples and the easiness with which they can be generated raise several security concerns with regard to deep learning systems, pushing researchers to develop suitable defense mechanisms. The use of networks adopting error-correcting output codes (ECOC) has recently been proposed to counter the creation of adversarial examples in a white-box setting. In this paper, we carry out an in-depth investigation of the adversarial robustness achieved by the ECOC approach. We do so by proposing a new adversarial attack specifically designed for multi-label classification architectures, like the ECOC-based one, and by applying two existing attacks. In contrast to previous findings, our analysis reveals that ECOC-based networks can be attacked quite easily by introducing a small adversarial perturbation. Moreover, the adversarial examples can be generated in such a way to achieve high probabilities for the predicted target class, hence making it difficult to use the prediction confidence to detect them. Our findings are proven by means of experimental results obtained on MNIST, CIFAR-10 and GTSRB classification tasks.

Introduction

Deep Neural Networks (DNNs) are notoriously fragile; small, imperceptible perturbations—known as adversarial examples—can trick them into misclassifying inputs. A recently proposed defense suggested using Error-Correcting Output Codes (ECOC) to mitigate this vulnerability.

In a standard network, the output is usually a "one-hot" vector. In an ECOC-based network, each class is assigned a long, unique binary codeword. To classify an image, the network must predict multiple bits correctly. The theory was that an attacker would find it much harder to flip multiple bits (logits) simultaneously than to flip just one, especially if those bits are generated by an ensemble of independent sub-networks. This paper investigates whether this "increased difficulty" actually holds up against a motivated attacker in a white-box setting.

Methods and Architecture

The ECOC architecture evaluated in this study uses a VGG-16 backbone where the final layers are split into an ensemble of branches. Each branch is responsible for predicting a specific bit of a Hadamard codeword.

Architecture of the ECOC-based network

The authors evaluate the ECOC defense by comparing three different attack strategies:

  1. Modified C&W Attack: An adaptation of the classic Carlini & Wagner attack, tweaked to handle the correlation-based probability function used in ECOC.
  2. LOTS Attack: Layerwise Origin-Target Synthesis, which attempts to make the deep-layer representation of a "source" image mimic a "target" image.
  3. The New Proposed Attack: This is the paper's core contribution. Instead of attacking the final probability distribution, this attack targets the individual logits ($z_i$) directly. It uses a Lagrangian optimization to minimize the distance between the current output and the target codeword bits ($t_i$). By focusing on the "bit-level" structure of the ECOC scheme, it bypasses the smoothing effects of the final activation layers.

Key Insights

  • The Multi-Label Weakness: While ECOC was intended to make attacks harder by requiring multiple bit-flips, training bits separately actually makes them easier to manipulate. An attacker can nudge individual bits toward a target codeword iteratively without significantly affecting others.
  • Confidence is Not a Safeguard: A major claim for ECOC was that adversarial examples would result in "low confidence" scores, allowing for detection. The authors proved this wrong by adding a "confidence margin" ($c$) to their attack, successfully generating adversarial images that the model classified into the wrong class with higher confidence than the original clean images.
  • Codeword Length is Insufficient: In tests using 32-bit codewords (GTSRB), the standard C&W attack struggled (42.3% success). However, the new bit-targeted attack achieved a 93.3% success rate, proving that longer codes do not guarantee security.

Experimental Results

The effectiveness of the attacks was tested on three standard benchmarks: MNIST, CIFAR-10, and GTSRB. The results show a clear "break" of the ECOC defense across all datasets.

Attack results on the CIFAR-10 dataset

On CIFAR-10, the proposed attack achieved a 98.6% Attack Success Rate (ASR) with a Peak Signal-to-Noise Ratio (PSNR) of 38.41 dB, meaning the noise is virtually invisible. On the more complex GTSRB task, the attack reached 93.3% ASR. Furthermore, the average probability assigned to the wrong target class on CIFAR-10 was pushed to 0.993 (99.3% confidence), effectively neutralizing confidence-based detection methods.

Conclusion

This study serves as a cautionary tale in the adversarial defense "arms race." It demonstrates that ECOC-based networks can be attacked quite easily by targeting the bit-level structure of the output codes. While the study is primarily white-box, it highlights fundamental weaknesses in the ECOC approach. Future work might involve exploring if ECOC could be combined with other defenses, like adversarial training, to provide a more meaningful security gain.

Summary updated

Abstract

The existence of adversarial examples and the easiness with which they can be generated raise several security concerns with regard to deep learning systems, pushing researchers to develop suitable defense mechanisms. The use of networks adopting error-correcting output codes (ECOC) has recently been proposed to counter the creation of adversarial examples in a white-box setting. In this paper, we carry out an in-depth investigation of the adversarial robustness achieved by the ECOC approach. We do so by proposing a new adversarial attack specifically designed for multi-label classification architectures, like the ECOC-based one, and by applying two existing attacks. In contrast to previous findings, our analysis reveals that ECOC-based networks can be attacked quite easily by introducing a small adversarial perturbation. Moreover, the adversarial examples can be generated in such a way to achieve high probabilities for the predicted target class, hence making it difficult to use the prediction confidence to detect them. Our findings are proven by means of experimental results obtained on MNIST, CIFAR-10 and GTSRB classification tasks.
BibTeX
@inproceedings{zhang2020challenging,
  author = {Bowen Zhang and Benedetta Tondi and Xixiang Lv and Mauro Barni},
  title = {{Challenging the adversarial robustness of DNNs based on error-correcting output codes}},
  booktitle = {Security},
  year = {2020},
  doi = {10.1155/2020/8882494},
}
Figure