Universal 3-Dimensional Perturbations for Black-Box Attacks on Video Recognition Systems

Shangyu Xie, Han Wang, Yu Kong, Yuan Hong

TL;DR

Researchers have developed U3D, the first black-box attack framework that generates universal 3-dimensional perturbations to subvert video recognition systems. By leveraging procedural noise and optimizing for spatio-temporal features, U3D creates adversarial videos that are highly effective across different AI models, nearly invisible to humans, and capable of bypassing state-of-the-art defenses.

What Problem Does This Paper Address?

Deep Learning models for video recognition are increasingly used in high-stakes environments, such as autonomous driving and CCTV anomaly detection. While adversarial attacks on static images are well-documented, attacking video is significantly harder.

Most existing video attacks simply apply 2D image perturbations to each frame individually. This approach has three major flaws:

  1. Lack of Temporal Consistency: Frame-by-frame changes often "flicker," making them easy for humans to see and for specialized defense algorithms (like AdvIT) to detect.
  2. Boundary Effects: If the perturbation isn't perfectly aligned with the video's start and end, the attack often fails.
  3. Computational Cost: Calculating a unique perturbation for every frame of a high-resolution video is too slow for real-time applications like attacking a live surveillance stream.

The authors propose U3D to solve these issues by creating a single, "universal" 3D noise volume that can be dropped onto any video at any time.

Methods

The core innovation of U3D lies in how it generates and optimizes noise:

  • Procedural Noise (Perlin & Gabor): Instead of trying to optimize millions of individual pixels, the authors use procedural noise functions. These functions use just a few parameters (like wavelength and orientation) to create smooth, natural-looking textures. This drastically reduces the "search space" for the attack.
  • Feature Space Optimization: Rather than just trying to flip the final label (e.g., changing "Running" to "Walking"), U3D optimizes the noise to maximize the distortion of features in the intermediate layers of a surrogate DNN. Because different models often learn similar features, this makes the attack highly "transferable" to models the attacker has never seen.
  • Particle Swarm Optimization (PSO): To find the perfect noise parameters, the authors use PSO—an evolutionary algorithm inspired by bird flocking—which is more efficient at finding global optima in complex spaces than traditional gradient descent for this specific task.
  • Temporal Transformation: To ensure the attack works regardless of when it starts, they integrate a "shifting" variable during training, making the perturbation robust against misalignment (the "boundary effect").

Datasets

The authors evaluated U3D using three major benchmarks:

  • HMDB51: 6,766 video clips of 51 human actions (used primarily to train the surrogate model).
  • UCF101: 13,320 clips of 101 action classes (used as a target for classification attacks).
  • UCF Crime: 1,900 long surveillance videos covering 13 anomalies like accidents and thefts (used to test attacks on real-world monitoring systems).

Key Insights and Ideas

  • 3D is Better than 2D: By treating the perturbation as a 3D volume (height x width x time), the attack respects the temporal nature of video. This consistency is why it bypasses defenses that look for "jitter" between frames.
  • Anomaly Detection is a Weak Link: One surprising finding was that anomaly detection systems were more vulnerable than standard action classifiers. The extra logic these systems use to score sequences makes them more sensitive to the specific types of feature distortion U3D produces.
  • Real-Time Viability: Because the noise is pre-calculated and "universal," it can be injected into a live stream with negligible latency (~0.015s per frame). This makes Man-in-the-Middle (MITM) attacks on unencrypted RTSP camera feeds a very real threat.

Key Results

  • High Success Rates: In black-box settings (where the attacker doesn't know the target model), U3D achieved success rates of 82.6% to 85.4% on UCF101 and over 90% on UCF Crime.
  • Superior Transferability: U3D outperformed the previous state-of-the-art (C-DUP). For example, a U3D attack designed on a C3D model had an 80%+ success rate when tested on totally different architectures like I3D or LRCN.
  • Human Stealth: In a user study with 87 participants, 96.4% of humans could not distinguish U3D-perturbed videos from the originals. In contrast, the baseline C-DUP attack was identified by humans 85.4% of the time.
  • Bypassing Defenses: U3D successfully bypassed AdvIT (a temporal consistency checker) with ease, whereas traditional frame-by-frame attacks were caught nearly 100% of the time.

Limitations and Future Work

While U3D is powerful, it is not invincible. The authors note that:

  • Adaptive Training: If a defender knows the specific U3D noise function, they can use "U3D-Adversarial Training" to make their models significantly more robust.
  • Encryption & Watermarking: The most practical defense against the "online" version of this attack is securing the communication channel (e.g., using encrypted RTSP) or using digital watermarking to detect stream tampering.
  • Future Models: The paper focused on current architectures like C3D and I3D. As video recognition moves toward newer "Transformer" based architectures, the transferability of these 3D perturbations will need further study.

Summary updated

Abstract

Widely deployed deep neural network (DNN) models have been proven to be vulnerable to adversarial perturbations in many applications (e.g., image, audio and text classifications). To date, there are only a few adversarial perturbations proposed to deviate the DNN models in video recognition systems by simply injecting 2D perturbations into video frames. However, such attacks may overly perturb the videos without learning the spatio-temporal features (across temporal frames), which are commonly extracted by DNN models for video recognition. To our best knowledge, we propose the first black-box attack framework that generates universal 3-dimensional (U3D) perturbations to subvert a variety of video recognition systems. U3D has many advantages, such as (1) as the transfer-based attack, U3D can universally attack multiple DNN models for video recognition without accessing to the target DNN model; (2) the high transferability of U3D makes such universal black-box attack easy-to-launch, which can be further enhanced by integrating queries over the target model when necessary; (3) U3D ensures human-imperceptibility; (4) U3D can bypass the existing state-of-the-art defense schemes; (5) U3D can be efficiently generated with a few pre-learned parameters, and then immediately injected to attack real-time DNN-based video recognition systems. We have conducted extensive experiments to evaluate U3D on multiple DNN models and three large-scale video datasets. The experimental results demonstrate its superiority and practicality.
BibTeX
@inproceedings{xie2022universal,
  author = {Shangyu Xie and Han Wang and Yu Kong and Yuan Hong},
  title = {{Universal 3-Dimensional Perturbations for Black-Box Attacks on Video Recognition Systems}},
  booktitle = {IEEE Symposium on Security and Privacy},
  year = {2022},
  url = {https://openreview.net/forum?id=zzUaU5u9jx},
}

Loading related content...

Figure