Skip to content
NeighborhoodCodingPublic

About

Description Real-time, model-agnostic adversarial defense for no-reference video quality assessment via inter-frame re-encoding. Code and data for the paper "ReCode".

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

3 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 

Repository files navigation

ReCode

Real-time, model-agnostic adversarial defense for no-reference video quality assessment (NR-VQA) via inter-frame re-encoding

Sungman Hong, AI Development Team, LIG Accuver. Paper under review (submitted to IEEE ICCE 2027).

ReCode system model

NR-VQA models that monitor streaming quality can be manipulated by imperceptible adversarial perturbations. We show that SecureVQA, to our knowledge the only defense designed for NR-VQA, is broken by an expectation-over-transformation (EOT) attack that averages the gradient over its test-time randomization. ReCode removes the perturbation from the input instead of hiding the gradient: the received video is re-encoded with an inter-frame codec (x264 on a server, the hardware HEVC encoder on a smartphone) at a fixed fraction ρ of its own bitrate before an unmodified quality model. The quantized residual discards most of the perturbation while the ranking of videos is preserved. No model access, no retraining.

Results on KoNViD-1k (1200 videos, SecureVQA)

Setting SRCC
SecureVQA, clean 0.868
SecureVQA, EOT attack (within its original budget) −0.220
SecureVQA + ReCode (ρ = 0.5), clean 0.835
SecureVQA + ReCode (ρ = 0.5), EOT attack 0.820
SecureVQA + ReCode (ρ = 0.5), adaptive attack (BPDA + EOT through the codec) 0.819
Frame-wise JPEG at the same quality loss (q = 30), EOT / adaptive attack 0.753 / 0.706

Re-encoding takes 0.17 s per 8-s 540p video with the fastest x264 preset.

Predicted versus true MOS on KoNViD-1k: clean, after the EOT attack, and after the EOT attack with ReCode

Planned release

  • Attack code: fixed-seed, EOT, and adaptive (BPDA + EOT) attacks on SecureVQA
  • ReCode defense wrapper for x264 and for Android MediaCodec (HEVC)
  • Per-video results for all 1200 KoNViD-1k videos: video name, MOS, and clean / attacked / defended scores for every x264 preset × rate ratio, frame-wise JPEG, intra-only coding, and TRADES fine-tuning
  • Android app that measures hardware HEVC re-encoding on the Samsung Galaxy S24 Ultra

To be updated.

Citation

@inproceedings{hong2027recode,
  title     = {ReCode: Real-Time, Model-Agnostic Adversarial Defense for Video Quality Assessment on Consumer Devices via Inter-Frame Re-encoding},
  author    = {Hong, Sungman},
  booktitle = {IEEE International Conference on Consumer Electronics (ICCE)},
  year      = {2027},
  note      = {under review}
}

About

Description Real-time, model-agnostic adversarial defense for no-reference video quality assessment via inter-frame re-encoding. Code and data for the paper "ReCode".

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors