Abstract
We propose a stealthy clean-label video backdoor attack against Deep Learning (DL)-based models aiming at detecting a particular class of spoofing attacks, namely video rebroadcast attacks. The injected backdoor does not affect spoofing detection in normal conditions, but induces a misclassification in the presence of a specific triggering signal. The proposed backdoor relies on a temporal trigger altering the average chrominance of the video sequence. The backdoor signal is designed by taking into account the peculiarities of the Human Visual System (HVS) to reduce the visibility of the trigger, thus increasing the stealthiness of the backdoor. To force the network to look at the presence of the trigger in the challenging clean-label scenario, we choose the poisoned samples used for the injection of the backdoor following a so-called Outlier Poisoning Strategy (OPS). According to OPS, the triggering signal is inserted in the training samples that the network finds more difficult to classify. The effectiveness of the proposed backdoor attack and its generality are validated experimentally on different datasets and anti-spoofing rebroadcast detection architectures.
Introduction
Deep Learning (DL) models are the backbone of modern face authentication systems, specifically for detecting "rebroadcast attacks"—where an adversary plays a video of an authorized user in front of a camera to gain unauthorized access. While these models are highly accurate, they are vulnerable to backdoor attacks. Most existing backdoor attacks are "dirty-label," meaning the attacker changes the training data labels. However, in real-world scenarios, data is often scrutinized by humans. This paper addresses the more challenging "clean-label" scenario: the attacker can only modify the video content, while the labels must remain correct. In this setting, a model typically ignores a backdoor trigger because it can learn to classify the "clean" features perfectly. The authors propose a trigger that is both human-stealthy and powerful enough to force a model to rely on it during training.
Methodology: The Clean-Label Pipeline
The proposed attack pipeline consists of three main components designed to overcome the clean-label incentive problem, where models tend to be "lazy" and ignore additional signals if standard features are sufficient.
- Temporal Chrominance Trigger: Instead of a static watermark, the trigger is a sinusoidal modulation of the average chrominance in the blue channel across video frames. The blue channel is chosen because the Human Visual System (HVS) is significantly less sensitive to blue light variations compared to red or green, making the trigger nearly imperceptible to humans.
- Outlier Poisoning Strategy (OPS): To force the model to learn the trigger, the authors target "outliers"—samples that are naturally difficult for the model to classify (those near the decision boundary). By poisoning these "hard" samples, the network is forced to look for additional signals (the trigger) to achieve high training accuracy.
- Ground-truth Feature Suppression (GFS): To further encourage the model to use the trigger, adversarial perturbations are used to subtly suppress the natural features that identify a video as "alive." This makes the backdoor trigger the most reliable feature for the model to latch onto.
Trigger Design and Stealthiness
The core of the attack is the temporal trigger. By modulating the chrominance rather than luminance, the attack achieves a high degree of stealthiness that traditional luminance-based attacks lack.
As shown in the visualization, the trigger's visibility increases with the modulation strength (Δ), but remains subtle for lower values. The researchers found they could use a very weak trigger during training to maintain stealth, but a stronger version of the same trigger at test time to ensure the backdoor "fires" reliably.
Experimental Results
The study utilized benchmarks like the Replay-attack and MSU-MFSD datasets, with VGGFace2 used for pre-training the face recognition component. Key results include:
- High Success Rate: Using the OPS-GFS strategy, the Attack Success Rate (ASR) reached nearly 100% in many configurations, even with a small fraction of poisoned data.
- Minimal Impact on Utility: Accuracy on clean, non-poisoned data remained high (above 95%), meaning the backdoor stays hidden during normal operation.
- Identity Preservation: The attack successfully bypassed the rebroadcast detector without affecting the subsequent face recognition module, which still identified users with 100% accuracy.
- Cross-Architecture Robustness: The attack remained effective even when the attacker used a different model architecture (e.g., AlexNet-LSTM) to generate the poison than the one used by the victim (e.g., ResNet18-LSTM).
Conclusion
This research highlights a significant vulnerability in video anti-spoofing systems. By exploiting human biological limitations and targeting difficult training samples, the proposed clean-label attack is both stealthy and highly effective. Future work will explore moving this attack into the physical domain, which would require modulating real-world lighting or displays to inject the chrominance trigger in real-time. The existence of such attacks underscores the urgent need for defensive techniques that can detect non-spatial, color-based anomalies in training data.
Summary updated