🕷️
SetupAI
繁
⚡Stalecollected in 2h

RaPA Pruning Boosts Transferable Attacks

RaPA Pruning Boosts Transferable Attacks
PostLinkedIn
⚡Read original on 雷峰网
#adversarial-attacks#transfer-attack#computer-vision#ai-securityraparapacvpr-2026

💡RaPA crushes defenses in cross-model attacks – critical for securing vision AI deployments.

⚡ 30-Second TL;DR

What Changed

Random pruning generates variant models for better generalization

Why It Matters

Raises stakes for AI vision model security, highlighting cross-arch vulnerabilities in black-box settings. Pushes need for robust defenses in autonomous systems.

What To Do Next

Reproduce RaPA from arXiv code to benchmark your vision model's adversarial robustness.

Who should care:Researchers & Academics

Key Points

  • •Random pruning generates variant models for better generalization
  • •45% avg success CNN-to-ViT attacks, +11.7-17.5% over SOTA
  • •88% success vs adv-trained models, tops all defenses tested
  • •Stronger gains with more compute; tested on ImageNet, 20+ models

🧠 Deep Insight

Background and context from public sources — not the original article. 8 sources cited.

🔑 Enhanced Key Takeaways

  • •RaPA's random pruning is mathematically equivalent to adding an importance regularization term that equalizes parameter contributions, reducing over-reliance on dominant parameters during adversarial optimization.[1]
  • •RaPA boosts ASRs by 14.6% on VGG16 and 20.7% on MBv2 when using Inception-v3 as surrogate, in addition to CNN-to-CNN transfers.[1]
  • •Performance scales with compute: with ResNet-50 surrogate, increasing iterations from 300 to 500 and forward-backward passes from 1 to 5 raises average ASR by 15.9%.[1]

🛠️ Technical Deep Dive

  • •At each optimization step, RaPA randomly prunes a subset of parameters in the surrogate model, generating multiple masked variants to update the adversarial example.[1]
  • •Parameter importance is measured by the change in loss if a parameter θ_i is removed, serving as a proxy for its contribution to attack effectiveness.[1]
  • •Theoretical basis: Random pruning encourages adversarial examples to depend less on specific parameter subsets, improving generalization across models.[1]

🔮 Future ImplicationsAI analysis grounded in cited sources

RaPA will raise the bar for black-box defense benchmarks in vision models
Its superior transferability from CNN to Transformer models and robustness against defenses like adversarial training will force reevaluation of existing SOTA defenses.
RaPA variants will appear in CVPR 2026 workshops on adversarial ML
Alignment with workshop themes on transferable attacks and agent robustness positions it for extension in competitions like jailbreak challenges at AdvML@CV.

⏳ Timeline

2025-04
RaPA paper first posted on arXiv (v1)
2025-11
RaPA paper updated to v2 on arXiv
2026-03
RaPA accepted to CVPR 2026
⚡Read original article on 雷峰网
📰

Weekly AI Recap

Read this week's curated digest of top AI events →

👉Related Updates

Same topic

Explore #adversarial-attacks

Same product

More on rapa

Same source

Latest from 雷峰网

DeepSeek Finally Unveils Vision Model

DeepSeek Finally Unveils Vision Model

cnBeta (Full RSS)•Aug 22

AI-curated news aggregator. All content rights belong to original publishers.
Original source: 雷峰网 ↗

This is a summary, not the original. Read the source, or get the weekly briefing.

Weekly AI briefing

One email a week. Unsubscribe anytime.