Unlearning-Enhanced Website Fingerprinting Attack: Against Backdoor Poisoning in Anonymous Networks

Yali Yuan, Kai Xu, Ruolin Ma, Yuchen ZhangarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces a "detoxification" framework for Website Fingerprinting (WF) attacks, making them resilient against hidden backdoor poisoning. By combining influence-based detection with machine unlearning via the Fisher Information Matrix, the authors created a model that can automatically identify and "forget" malicious data, maintaining ~80% accuracy in adversarial settings where traditional models completely fail.

What Problem Does This Paper Address?

Website Fingerprinting (WF) is a technique used by researchers and authorities to identify which websites a user is visiting over anonymous networks like Tor by analyzing traffic patterns (packet sizes, timing, and direction). While deep learning has made WF attacks highly effective, these models have a "glass jaw": backdoor poisoning.

In a backdoor attack, a malicious actor (like a website admin) injects a tiny, specific "trigger" into the traffic data. If this poisoned data ends up in the training set, the model learns to associate that trigger with a wrong label. To a human observer, the model looks perfect because it performs normally on clean data. However, when the attacker activates the trigger in the real world, the model is "tricked" into misclassifying the traffic, effectively paralyzing the surveillance tool.

The challenge is that detecting these poisoned points manually is like finding a needle in a haystack, and retraining a massive model from scratch every time a poison is suspected is computationally expensive.

Methods

The authors propose a two-stage solution: Detection and Unlearning.

  1. Few-Shot Poisoned Point Detection: Instead of scanning the whole dataset, the method starts with just a few known "poisoned" test points (samples that cause the model to fail). They use influence scores to see which training samples contributed most to those failures.

  2. The Novel Twist: They apply random transformations (inserting, splitting, or merging traffic units) to these test points. Clean data is robust to these shifts, but backdoor triggers are brittle. By comparing influence scores before and after these transformations, they can accurately flag poisoned training samples.

  3. Parameter-based Unlearning: Once the poisoned samples are identified, the model needs to "forget" them without losing its ability to recognize clean websites. The authors use the Fisher Information Matrix (FIM) to calculate the importance of every model parameter.

  4. They identify "specialized" parameters—those that are highly sensitive to the poisoned data but have little impact on clean data.

  5. Instead of retraining, they perform a "surgical strike" by suppressing these specific parameters. This is known as approximate unlearning, which is significantly faster than starting from scratch.

Datasets

The study utilized two large-scale public datasets from the WFlib library, representing the two standard scenarios in the field:

  • tor_100w_2500tr (Closed-World): Contains ~250,000 fingerprints from 100 different websites. In this scenario, the attacker assumes the user only visits these 100 sites.
  • tor_open_200w_2000tr (Open-World): Contains ~400,000 fingerprints from 200 monitored sites plus "unknown" sites. This is a more realistic, "noisier" environment.

Key Insights and Ideas

  • Triggers are Brittle: The most clever insight is using data augmentation (trace enhancement) as a diagnostic tool. Because backdoor triggers are often artificial and "overfit" to specific patterns, they don't survive random network trace manipulations as well as natural website features do.
  • Surgical Forgetting: The authors lean on the theory that deep networks store specific "memorized" examples in later layers. By targeting these specialized parameters using FIM, they can remove the backdoor's influence without damaging the model's general knowledge.
  • Efficiency is King: By opting for parameter adjustment over full retraining, the framework achieves "detoxification" while actually running faster than traditional WF models.

Key Results

The proposed method was tested against state-of-the-art WF models like Deep Fingerprinting (DF) and Triplet Fingerprinting (TF):

  • Resilience: In poisoned scenarios where traditional models' accuracy dropped to near 0% (on poisoned test sets), the authors' model maintained a stable accuracy of approximately 80%.
  • Clean Performance: The "unlearning" process didn't break the model; it maintained roughly 95% accuracy on clean datasets.
  • Speed: The framework achieved a 2-3x speedup in runtime efficiency compared to baseline methods, making it much more practical for real-world deployment.

Limitations and Future Work

While the results are promising, the authors acknowledge several areas for improvement:

  • Generality: The current unlearning algorithm is specifically tuned for backdoor attacks. Future work could explore integrating it with differential privacy to handle a broader range of adversarial threats.
  • Network Dynamics: Real-world Tor traffic is highly volatile. The model needs better robustness against natural network fluctuations and frequent website content updates.
  • Scalability: While faster than retraining, the authors note that further engineering is needed to handle truly massive, distributed datasets in real-time production environments.

Summary updated

Abstract

Website Fingerprinting (WF) is an effective tool for regulating and governing the dark web. However, its performance can be significantly degraded by backdoor poisoning attacks in practical deployments. This paper aims to address the problem of hidden backdoor poisoning attacks faced by Website Fingerprinting attack, and designs a feasible mothed that integrates unlearning technology to realize detection of automatic poisoned points and complete removal of its destructive effects, requiring only a small number of known poisoned test points. Taking Tor onion routing as an example, our method evaluates the influence value of each training sample on these known poisoned test points as the basis for judgment. We optimize the use of influence scores to identify poisoned samples within the training dataset. Furthermore, by quantifying the difference between the contribution of model parameters on the taining data and the clean data, the target parameters are dynamically adjusted to eliminate the impact of the backdoor attacks. Experiments on public datasets under the assumptions of closed-world (CW) and open-world (OW) verify the effectiveness of the proposed method. In complex scenes containing both clean website fingerprinting features and backdoor triggers, the accuracy of the model on the poisoned dataset and the test dataset is stable at about 80%, significantly outperforming the traditional WF attack models. In addition, the proposed method achieves a 2-3 times speedup in runtime efficiency compared to baseline methods. By incorporating machine unlearning, we realize a WF attack model that exhibits enhanced resistance to backdoor poisoning and faster execution speeds in adversarial settings.
BibTeX
@misc{yuan2025unlearning,
  author = {Yali Yuan and Kai Xu and Ruolin Ma and Yuchen Zhang},
  title = {{Unlearning-Enhanced Website Fingerprinting Attack: Against Backdoor Poisoning in Anonymous Networks}},
  year = {2025},
  eprint = {2506.13563},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2506.13563},
  howpublished = {arXiv},
}
Figure