Discerning the Chaos: Detecting Adversarial Perturbations while Disentangling Intentional from Unintentional Noises

Anubhooti Jain, Susim Mukul Roy, Kwanit Gupta, Mayank Vatsa, Richa Singh

Introduction

Deep neural networks (DNNs) are remarkably effective but notoriously "brittle." They are susceptible to adversarial attacks—subtle, intentional perturbations designed to deceive a classifier into making an incorrect prediction (e.g., misclassifying a "Male" face as "Female"). Simultaneously, real-world images frequently encounter unintentional noise, such as Gaussian blur, sensor artifacts, or compression.

Overview of intentional vs. unintentional image perturbations.

The primary challenge in securing these systems is twofold: detecting when an image has been modified and, crucially, disentangling the intent behind the modification. Distinguishing between a "bad photo" (unintentional noise) and a malicious "hack" (intentional attack) allows for more nuanced defensive responses. This paper introduces CIAI (Class-Independent Adversarial Intent), a detector network designed to address this gap by focusing on the statistical signatures of noise intent.

Methodology

The CIAI network utilizes a Vision Transformer (ViT) backbone, chosen for its superior generalization capabilities under distribution shifts compared to standard Convolutional Neural Networks (CNNs). The framework employs a two-stage training process focused on feature clustering and classification.

Two-Stage Training

  1. Feature Clustering: The network learns to map images into an embedding space where original images, various adversarial attacks, and different unintentional noises form distinct clusters. This is achieved through a combination of Maximum Mean Discrepancy (MMD), which compares distributions of image sets, and Center Loss, which pulls embeddings of the same noise type toward a central point while pushing different types apart.
  2. Classification: Once the encoder is trained and frozen, simple linear layers are used to perform the final classification into three categories: Original, Intentional, or Unintentional.

Loss Function ($L_{det}$)

The detection loss is a weighted sum of three components designed to maximize separation:

  • $L_1$: Separates original images from all modified versions.
  • $L_2$: Distinguishes between intentional attacks and unintentional noises.
  • $L_3$: Creates separation within subgroups, such as distinguishing between FGSM and PGD attacks.

Experimental Setup

The model was evaluated across several prominent benchmarks to test its robustness and generalization:

  • Face Datasets: CelebA, CelebA-HQ, LFW (Labeled Faces in the Wild), and AgeDB, primarily for gender attribute prediction.
  • General Object Recognition: CIFAR-10 and its corrupted counterpart, CIFAR-10-C.
  • Perturbation Types: The model was trained on "seen" modifications (FGSM, PGD, Gaussian, Salt & Pepper) and tested on "unseen" ones (BIM, DeepFool, CW, Speckle noise).

Key Results and Insights

CIAI demonstrates high detection accuracy and impressive generalization to novel threats. On the CelebA dataset, the model achieved nearly 99.9% accuracy in detecting seen attacks. More importantly, it maintained high performance on unseen attacks, detecting the BIM attack with 99.92% accuracy and the CW attack with 84.2% accuracy on the LFW dataset, despite never encountering them during training.

Class Independence and Noise Signatures

A key advantage of CIAI is its class-independent nature; it detects perturbations without needing to know the underlying image content (e.g., whether the subject is a person or an object). Analysis via t-SNE visualizations confirms that different modifications occupy specific regions in the latent space. Even unseen attacks tend to cluster near seen attacks of the same family, enabling effective detection of zero-day adversarial threats.

Conclusion

By identifying the intent of image noise, CIAI provides a sophisticated security layer for DNNs. While the current focus has been on facial attributes like gender, future work could expand this to complex multi-attribute tasks, facial recognition, or physical-world adversarial attacks. Despite the higher computational cost of ViTs, the ability to disentangle malicious intent from benign corruption represents a significant step forward in building robust and trustworthy AI systems.

Summary updated

BibTeX
@inproceedings{jain2024discerning,
  author = {Anubhooti Jain and Susim Mukul Roy and Kwanit Gupta and Mayank Vatsa and Richa Singh},
  title = {{Discerning the Chaos: Detecting Adversarial Perturbations while Disentangling Intentional from Unintentional Noises}},
  booktitle = {IJCB},
  year = {2024},
  doi = {10.1109/ijcb62174.2024.10744526},
}
Figure