SURA detects federated-learning backdoors with 10 clean images
It reuses one small clean proxy across repeated loss-ascent probes and restores batch-norm state only at readout.
Stories per day · 30 days
1 RSS
Relatedbackdoor1bagel1batch-normalization1federated-learning1janus-pro1mixture-of-experts1model-poisoning1model-safety1multimodal-models1wmdp1
11 stories, newest first · one card per story
Stories tagged #unlearning
UniForget enables joint concept unlearning in multimodal models
UniForget achieves 80.0% joint forgetting on BAGEL-7B-MoT and 92.7% on Janus-Pro-7B, outperforming single-pathway baselines while preserving other capabilities.
REX moves hazardous knowledge into removable experts in MoE models
Excising those experts lowers WMDP-Bio by 48.4 points on Qwen3.5-35B-A3B while MMLU changes by less than 3 points.
Certified Mechanistic Edits: Behavioral Guarantees for Skill Removal and Preservation
Researchers certify that mechanistic edits provably remove one neural network skill while preserving another, replacing test-only validation with formal behavioral guarantees.
Faithful Dual-constrained Erasure for Robust LLM Safety Alignment
FDCU uses dual-constrained projection to erase hazardous LLM knowledge and resist retraining attacks that revive unlearned behavior.
What Does It Mean to Forget a Person? Individual-Level Unlearning in Vision-Language Models
IDUnlearn-Bench finds vision-language models can drop a person's attributes yet still identify and reconstruct them.
The Tokens Remember: When Tokenization Bypasses Knowledge Editing and Unlearning
Toketive shows alternative tokenizations can bypass LLM knowledge edits and recover suppressed answers.
A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
PIR reads internal states to separate concealed model knowledge from genuine ignorance across eight language models.
A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
PIR reads internal model states to detect concealed knowledge, distinguishing sandbagging from genuine ignorance across Gemma, Qwen, Llama, Mistral and Phi models.
Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents
Researchers propose provenance-guided selective replay letting LLM agents forget revoked information without restarts, matching full reset behavior.
Client-Side Probing of Deleted Ridge Statistics in Federated Unlearning
Researchers show malicious federated learning clients can probe broadcast classifiers to recover deleted samples, exposing exact label leakage on MNIST and CIFAR-10.