SNEAK: Synonymous Sentences-Aware Adversarial Attack on Natural Language Video Localization

Wenbo Gou, Wen Shi, Jian Lou, Lijie Huang, Pan Zhou, Ruixuan Li

Introduction

Natural Language Video Localization (NLVL) is a critical task in computer vision that involves identifying the start and end timestamps of specific events within untrimmed videos based on natural language queries. As these models find applications in security-sensitive domains such as surveillance and crime detection, understanding their robustness against adversarial attacks is paramount. While adversarial research has matured in single-modality domains like image classification, the cross-modal vulnerabilities of NLVL models—specifically how the interplay between video and text can be exploited—remain largely unexplored.

The SNEAK Framework

The authors introduce SNEAK (Synonymous seNtences-awarE Adversarial AttacK), a novel framework designed to expose the fragility of NLVL models. Unlike traditional attacks that target a single modality or a specific query, SNEAK leverages the semantic relationship between synonymous text queries to create more robust and devastating perturbations.

Overview of attack modalities: (a) text-only, (b) video-only, and (c) joint video and query attack.

SNEAK comprises three primary innovations:

  1. Synonymous Sentence Awareness: Using a "max-min" optimization strategy, SNEAK generates video perturbations that maximize model loss across a set of synonymous queries. This ensures the attack remains effective even if the user rephrases their search (e.g., "cutting a cake" vs. "slicing a dessert").
  2. Gradient Splitting Optimization: To handle the high computational cost of attacking massive NLVL models, the authors split the gradient calculation. They first optimize perturbations for extracted video features and then use a gradient-descent fitting method to map these changes back to the original video pixels.
  3. Frame Importance-Guided Pruning (PSA): To maintain stealthiness, the Frame Importance-Guided Pruning (PSA) mechanism identifies and perturbs only the most influential frames, minimizing the visual footprint of the attack.

Experimental Results

The SNEAK attack was evaluated on the TACoS and Charades-STA benchmarks using the VSLNet model. The results demonstrate a significant vulnerability in state-of-the-art NLVL architectures.

Quantitative comparison of mIoU across different queries under various attack scenarios, showing the effectiveness of SNEAK and SNEAK+PSA.

On the TACoS dataset, the baseline model's Mean Intersection over Union (mIoU) plummeted from 24.11% to approximately 3-4% under a full SNEAK attack. Notably, the PSA method allowed for pruning up to 75% of the perturbations while maintaining high attack effectiveness, suggesting that NLVL models are overly reliant on a small subset of "important" frames.

Conclusion and Future Work

SNEAK reveals that joint cross-modal attacks are significantly more effective than single-modality approaches. By exploiting linguistic synonyms, the attack uncovers deeper weaknesses in how models associate visual features with semantic meaning. While adversarial training offers some defense, it comes with a trade-off in clean-data accuracy. Future research will focus on extending these attacks to "black-box" scenarios and investigating whether these vulnerabilities are inherent to the NLVL task itself.

Summary updated

Abstract

Natural language video localization (NLVL) is an important task in the vision-language understanding area, which calls for an in-depth understanding of not only computer vision and natural language side alone, but more importantly the interplay between both sides. Adversarial vulnerability has been well-recognized as a critical security issue of deep neural network models, which requires prudent investigation. Despite its extensive yet separated studies in video and language tasks, current understanding of the adversarial robustness in vision-language joint tasks like NLVL is less developed. This paper therefore aims to comprehensively investigate the adversarial robustness of NLVL models by examining three facets of vulnerabilities from both attack and defense aspects. To achieve the attack goal, we propose a new adversarial attack paradigm called synonymous sentences-aware adversarial attack on NLVL (SNEAK), which captures the cross-modality interplay between the vision and language sides.
BibTeX
@misc{gou2021sneak,
  author = {Wenbo Gou and Wen Shi and Jian Lou and Lijie Huang and Pan Zhou and Ruixuan Li},
  title = {{SNEAK: Synonymous Sentences-Aware Adversarial Attack on Natural Language Video Localization}},
  year = {2021},
  eprint = {2112.04154},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.04154},
}
Figure