Skip to main content
Back

ViPro Forces Any Video to Top Search Results Across Multiple Queries

A new adversarial technique pushes attacker videos to the top of retrieval results across multiple queries at once, exposing a fresh vulnerability in text-to-video search.

Pro

  • ViPro is the first adversarial attack designed to promote videos in text-to-video retrieval, not suppress them.
  • The attack pushes a video into the intersection of retrieval boundaries for multiple queries at once.
  • Modality Refinement (MoRe) groups frames into clips and uses semantic weighting to resolve conflicting gradients.
  • Beats prior baselines by 30/10/4% in white/grey/black-box settings across Singularity, DRL, Cap4Video.
  • Survives JPEG compression and Temporal Shuffling defenses better than adapted image-domain attacks.
  • Code is available at github.com/michaeltian108/ViPro; commercial platforms remain harder due to non-visual signals.

Adversarial attacks on retrieval systems have historically focused on knocking legitimate results down the list. A new paper inverts that goal: rather than suppressing videos, it promotes them, forcing a chosen video to appear as the top match for multiple text queries simultaneously. The attack, called ViPro (Video Promotion attack), targets modern text-to-video retrieval (T2VR) systems with perturbations subtle enough to survive human review.

Why promotion is harder than suppression

Text-to-video retrieval has grown quickly on the back of vision-language models, but its robustness remains largely unexamined. Prior work has pushed videos away from queries to suppress their ranks, while pulling videos toward selected queries has been left unexplored. The economic incentive is obvious: promotion can yield more views, clicks, ad revenue, and reach for misinformation.

Geometrically, the two problems are not symmetric. The retrieval boundary is the region in which videos can be retrieved by a query. Suppression only has to escape one sphere. Promotion has to land inside the intersection of many such spheres, and that intersection is small.

Diagram comparing video suppression and video promotion, showing retrieval boundaries and target overlap area

The core loss trick

A naive approach would minimize negative similarity between the video and each target query. The authors show this fails in multi-target settings because the gradients treat all queries identically, ignoring that some are already close and some are far. ViPro instead uses an exponential loss over the sum of cross-modal similarities. That produces larger gradients for queries the video is currently far from, and smaller gradients for queries it has already reached, preventing easy targets from dominating optimization.