Scholarlink

<p>Deep video recognition systems excel at spatio–temporal understanding yet remain vulnerable to adversarial manipulation, and most prior video attacks perturb all or most frames, increasing artifacts, query cost, and detectability. We introduce a time-constrained adversarial attack that enforces temporal sparsity via a frame-level mask which activates perturbations on exactly <i>K</i> frames while keeping all remaining frames identical to the original through a final mask-consistent identity projection step. Within the masked frames, our method employs a score-based update that stochastically estimates and rectifies an ascent direction and then applies a projection under an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\ell _\infty \)</EquationSource> </InlineEquation> budget; an optional temporal TV–L1 term smooths perturbations across adjacent attacked frames. The approach supports both targeted and untargeted goals, and we formalize the problem, detail the optimization and masking strategies, and analyze trade-offs among temporal sparsity, perturbation budget, and query allowance. We implement the attack on an I3D classifier with Kinetics-400 and discuss implications for temporally aware defenses and responsible disclosure. Experimental results show that the proposed attack achieves a 96.3% success rate in the targeted setting at 5,000 iterations and a 99.7% success rate in the untargeted setting at 1,000 iterations, while preserving high perceptual similarity.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Time-constrained adversarial attacks for video recognition models: temporally sparse but effective perturbations

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Joo Bon Maeng,
  • Hyun Kwon

摘要

Deep video recognition systems excel at spatio–temporal understanding yet remain vulnerable to adversarial manipulation, and most prior video attacks perturb all or most frames, increasing artifacts, query cost, and detectability. We introduce a time-constrained adversarial attack that enforces temporal sparsity via a frame-level mask which activates perturbations on exactly K frames while keeping all remaining frames identical to the original through a final mask-consistent identity projection step. Within the masked frames, our method employs a score-based update that stochastically estimates and rectifies an ascent direction and then applies a projection under an \(\ell _\infty \) budget; an optional temporal TV–L1 term smooths perturbations across adjacent attacked frames. The approach supports both targeted and untargeted goals, and we formalize the problem, detail the optimization and masking strategies, and analyze trade-offs among temporal sparsity, perturbation budget, and query allowance. We implement the attack on an I3D classifier with Kinetics-400 and discuss implications for temporally aware defenses and responsible disclosure. Experimental results show that the proposed attack achieves a 96.3% success rate in the targeted setting at 5,000 iterations and a 99.7% success rate in the untargeted setting at 1,000 iterations, while preserving high perceptual similarity.