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

Shangyu Xie, Han Wang, Yu Kong, Yuan Hong

Abstract

Widely deployed deep neural network (DNN) models have been proven to be vulnerable to adversarial perturbations in many applications. 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. 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.

Introduction

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 due to the temporal dimension. 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 applied to any video at any time.

Methodology

The core innovation of U3D lies in how it generates and optimizes noise to exploit spatio-temporal features.

The U3D framework overview, showing the generator (a) and the injection process (b).

Procedural Noise and Optimization

Instead of trying to optimize millions of individual pixels, the authors use procedural noise functions (Perlin and Gabor). These functions use just a few parameters (like wavelength and orientation) to create smooth, natural-looking textures, drastically reducing the search space for the attack. To find the optimal noise parameters, the authors employ Particle Swarm Optimization (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.

Feature Space Optimization and Temporal Robustness

Rather than just trying to flip the final label, 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. 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").

Experimental Evaluation

The authors evaluated U3D using three major benchmarks: HMDB51 (for training the surrogate model), UCF101 (for action classification attacks), and UCF Crime (for real-world surveillance anomaly detection).

Key Results

U3D demonstrated significant superiority over existing methods:

  • High Success Rates: In black-box settings, 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 architectures like I3D or LRCN.
  • Human Stealth: In a user study with 87 participants, 96.4% could not distinguish U3D-perturbed videos from the originals.
  • Bypassing Defenses: U3D successfully bypassed AdvIT (a temporal consistency checker), whereas traditional frame-by-frame attacks were caught nearly 100% of the time.
Success rates of U3D attacks across different DNN models (I3D, DN, LRCN, TSN) as the number of queries increases.

Discussion and Conclusion

The research highlights that treating perturbations as a 3D volume respects the temporal nature of video, allowing attacks to bypass defenses looking for inter-frame "jitter." Furthermore, anomaly detection systems were found to be particularly vulnerable due to their sensitivity to feature distortions.

Because the noise is pre-calculated and "universal," it can be injected into a live stream with negligible latency (~0.015s per frame), making Man-in-the-Middle (MITM) attacks on unencrypted RTSP camera feeds a credible threat. Future defenses may need to focus on adaptive training, encrypted communication channels, or digital watermarking to detect stream tampering.

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},
}
Figure