Cascading Adversarial Bias from Injection to Distillation in Language Models

• URL: http://arxiv.org/abs/2505.24842v1
• Date: Fri, 30 May 2025 17:41:58 GMT
• Title: Cascading Adversarial Bias from Injection to Distillation in Language Models
• Authors: Harsh Chaudhari, Jamie Hayes, Matthew Jagielski, Ilia Shumailov, Milad Nasr, Alina Oprea,
• Abstract summary: This paper investigates vulnerability of distilled models to adversarial injection of biased content during training.<n>We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning.<n>We propose practical design principles for building effective adversarial bias mitigation strategies.
• Score: 38.38782840888917
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Model distillation has become essential for creating smaller, deployable language models that retain larger system capabilities. However, widespread deployment raises concerns about resilience to adversarial manipulation. This paper investigates vulnerability of distilled models to adversarial injection of biased content during training. We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning, which propagates to student models and becomes significantly amplified. We propose two propagation modes: Untargeted Propagation, where bias affects multiple tasks, and Targeted Propagation, focusing on specific tasks while maintaining normal behavior elsewhere. With only 25 poisoned samples (0.25% poisoning rate), student models generate biased responses 76.9% of the time in targeted scenarios - higher than 69.4% in teacher models. For untargeted propagation, adversarial bias appears 6x-29x more frequently in student models on unseen tasks. We validate findings across six bias types (targeted advertisements, phishing links, narrative manipulations, insecure coding practices), various distillation methods, and different modalities spanning text and code generation. Our evaluation reveals shortcomings in current defenses - perplexity filtering, bias detection systems, and LLM-based autorater frameworks - against these attacks. Results expose significant security vulnerabilities in distilled models, highlighting need for specialized safeguards. We propose practical design principles for building effective adversarial bias mitigation strategies.

Related papers

• Exploiting Edge Features for Transferable Adversarial Attacks in Distributed Machine Learning [54.26807397329468]
  This work explores a previously overlooked vulnerability in distributed deep learning systems.<n>An adversary who intercepts the intermediate features transmitted between them can still pose a serious threat.<n>We propose an exploitation strategy specifically designed for distributed settings.
  arXiv  Detail & Related papers  (2025-07-09T20:09:00Z)
• DUMB and DUMBer: Is Adversarial Training Worth It in the Real World? [15.469010487781931]
  Adversarial examples are small and often imperceptible perturbations crafted to fool machine learning models.<n>Evasion attacks, a form of adversarial attack where input is modified at test time to cause misclassification, are particularly insidious due to their transferability.<n>We introduce DUMBer, an attack framework built on the foundation of the DUMB methodology to evaluate the resilience of adversarially trained models.
  arXiv  Detail & Related papers  (2025-06-23T11:16:21Z)
• Defending Against Frequency-Based Attacks with Diffusion Models [0.23020018305241333]
  Diffusion models have proven highly effective for noise purification, not only in countering pixel-wise adversarial perturbations but also in addressing non-adversarial data shifts.<n>Our findings highlight its effectiveness in handling diverse distortion patterns across low- to high-frequency regions.
  arXiv  Detail & Related papers  (2025-04-15T09:57:17Z)
• FreqFed: A Frequency Analysis-Based Approach for Mitigating Poisoning Attacks in Federated Learning [98.43475653490219]
  Federated learning (FL) is susceptible to poisoning attacks. FreqFed is a novel aggregation mechanism that transforms the model updates into the frequency domain. We demonstrate that FreqFed can mitigate poisoning attacks effectively with a negligible impact on the utility of the aggregated model.
  arXiv  Detail & Related papers  (2023-12-07T16:56:24Z)
• Isolation and Induction: Training Robust Deep Neural Networks against Model Stealing Attacks [51.51023951695014]
  Existing model stealing defenses add deceptive perturbations to the victim's posterior probabilities to mislead the attackers. This paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses. In contrast to adding perturbations over model predictions that harm the benign accuracy, we train models to produce uninformative outputs against stealing queries.
  arXiv  Detail & Related papers  (2023-08-02T05:54:01Z)
• AdvDiff: Generating Unrestricted Adversarial Examples using Diffusion Models [7.406040859734522]
  Unrestricted adversarial attacks present a serious threat to deep learning models and adversarial defense techniques. Previous attack methods often directly inject Projected Gradient Descent (PGD) gradients into the sampling of generative models. We propose a new method, called AdvDiff, to generate unrestricted adversarial examples with diffusion models.
  arXiv  Detail & Related papers  (2023-07-24T03:10:02Z)
• Exploring Model Dynamics for Accumulative Poisoning Discovery [62.08553134316483]
  We propose a novel information measure, namely, Memorization Discrepancy, to explore the defense via the model-level information. By implicitly transferring the changes in the data manipulation to that in the model outputs, Memorization Discrepancy can discover the imperceptible poison samples. We thoroughly explore its properties and propose Discrepancy-aware Sample Correction (DSC) to defend against accumulative poisoning attacks.
  arXiv  Detail & Related papers  (2023-06-06T14:45:24Z)
• Defending against the Label-flipping Attack in Federated Learning [5.769445676575767]
  Federated learning (FL) provides autonomy and privacy by design to participating peers. The label-flipping (LF) attack is a targeted poisoning attack where the attackers poison their training data by flipping the labels of some examples. We propose a novel defense that first dynamically extracts those gradients from the peers' local updates.
  arXiv  Detail & Related papers  (2022-07-05T12:02:54Z)
• Learning to Attack: Towards Textual Adversarial Attacking in Real-world Situations [81.82518920087175]
  Adversarial attacking aims to fool deep neural networks with adversarial examples. We propose a reinforcement learning based attack model, which can learn from attack history and launch attacks more efficiently.
  arXiv  Detail & Related papers  (2020-09-19T09:12:24Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.