Search: “adversarial perturbation”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

What Makes Adversarial Examples Transfer Across Deepfake Detectors?

A controlled study of 60 deepfake detectors shows adversarial example transfer depends heavily on source-target compatibility, with source averaging understating vulnerability.

The study evaluates adversarial example transferability across 60 deepfake detectors spanning six backbones, two pretraining regimes, and five training-data configurations, using AutoAttack (AA) and Carlini-Wagner with Expectation over Transformation (CW-EOT). Transfer rises sharply when source and target share an exact backbone, architecture family, pretraining regime, or training data, with the dominant factor depending on the attack. Mean attack success rate is 7.21% under AA and 19.52% under CW-EOT for single sources, while a multi-source oracle reaches 64.48% after excluding exact matches, showing source averaging can substantially understate target vulnerability. The authors release 240,000 adversarially perturbed images, pairwise transfer results, detector configurations, and evaluation code.

arXiv cs.CR·Sep 9, 10:28 UTC · 7d agoAI safety & security

40

Robust Policy Optimization via Adversarial Importance Sampling

Adversarial Importance Sampling estimates worst-case RL returns without extra interactions; authors also release the advrl PyTorch library.

The paper introduces Advis, which uses importance sampling over trajectories from standard training to estimate and optimize verifiable worst-case returns, requiring no additional environment interactions or auxiliary networks. It also releases advrl, a modular PyTorch library of single-file robustness methods and adversarial attacks for reproducible evaluation. The authors show adversarial hyperparameters do not transfer across agents, so they evaluate with 6-14x more attacker configurations than prior work. Effectiveness is demonstrated on continuous control environments.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:40 UTC · 5d agoAI research1

20

RAPID: A Real-Time Defense Against Unauthorized Model Distillation for Text-to-Image Services

RAPID embeds defensive perturbations in a T2I model's shared VAE decoder to block unauthorized black-box distillation in real time.

The paper defends text-to-image services against model theft via black-box output-based distillation, where adversaries collect prompt-image pairs to train substitute models. RAPID integrates defensive perturbations into the shared VAE decoder using self-referenced latent maximization plus reconstruction-guided color regularization, avoiding costly sample-wise online optimization. Across four T2I models and four datasets versus five baselines, it consistently degrades substitute-model generation quality while preserving visual fidelity.

arXiv cs.CR·Sep 14, 16:11 UTC · 2d agoAI safety & security

35

A Unified and Constrained View of Regularization-Based Robust Reinforcement Learning

Paper unifies regularization-based robust RL methods via new performance-gap upper bounds and jointly learned Lagrange multipliers.

The authors derive new upper bounds on the gap between nominal and worst-case deep RL policies, each expressible as an existing regularization objective plus a KL-divergence penalty. Robust training is reformulated as constrained optimization, where prior methods correspond to a fixed Lagrange multiplier. The multiplier is instead updated jointly with the policy, auto-tuning the regularization weight. Adversarial evaluations across several continuous control tasks validate the theory.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 16:45 UTC · 5d agoAI research1

20

Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning

Researchers analyze why Preventative Steering protects LLMs against malicious fine-tuning, finding active adaptation drives protection, and propose Progressive Intensity Scheduling.

The paper studies Preventative Steering, a training-time defense that injects undesirable-trait persona vectors during adversarial fine-tuning and removes them at evaluation time. Temporal analysis shows protection emerges from an early compensatory adaptation phase followed by a steady-state phase, with attention output projections acting as the dominant residual-write route for defensive updates. Intervention Delta Preservation experiments show that preserving or reinjecting weight offsets fails to maintain protection, indicating reliance on active adaptation rather than a static defense. The proposed Progressive Intensity Scheduling improves safety robustness on Qwen2.5 and Gemma-3 while reducing harmful trait expression.

arXiv cs.CR·Sep 9, 13:19 UTC · 7d agoAI safety & security1

32

Do Input-Level Defenses Transfer to Observation-Level Attacks on VideoLLMs?

A systematic study shows input-level adversarial defenses provide inconsistent, often near-zero protection against observation-level attacks on video LLMs.

Researchers introduce DefTEval, a controlled framework testing eleven input-level defenses against five attack types across five video LLMs. Harmful-content detection rates are frequently near zero, and defenses fail even when attacks embed harmful signals in every sampled frame. Token compression discards localized safety features and modality fusion down-weights weakened visual signals, with defense outcomes dominated by model architecture rather than the defense method.

arXiv cs.CR·Sep 8, 06:59 UTC · 8d agoAI safety & security

40

Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety

Unit 42 research shows LLM safety refusals concentrate in a thin neural layer, motivating external, multi-layered AI security controls.

Palo Alto Networks Unit 42 introduces Perturbation Probing, a diagnostic technique for measuring the fragility of LLM safety mechanisms. The research finds that safety refusal behavior is localized within a thin neural layer, implying small perturbations can undermine built-in refusals. The authors argue this motivates external, multi-layered security defenses on top of model-internal safety training.

Palo Alto Unit 42·Aug 28, 22:00 UTC · 19d agoAI safety & security

45

Adversarial Testing of Automated Program Repair Agents for Security Vulnerabilities

SWEADV benchmark shows adversarial issue descriptions make LLM program-repair agents write insecure fixes in 51.7% of cases, evading most detection tools.

Researchers built SWEADV, a benchmark of 750 adversarial issue descriptions derived from 150 SWE-bench Verified repair tasks, covering command execution, deserialization, path traversal, denial of service, and weak hashing attack types. Tested on mini_swe agents backed by GPT-5-Mini, MiniMax-M2.5, and DeepSeek-R, adversarial descriptions induced malicious behavior with successful repair in 51.7% of cases. Detection was weak: LLM-as-judge pre-repair screening reached only 62.3% accuracy, and post-repair detection via static analysis and LLM-as-judge achieved just 39.4% and 55.4%.

arXiv cs.CR·Sep 14, 17:53 UTC · 2d agoAI safety & security2

52

On Identifying Adversarial Intent Injection in AI-Native 6G Networks

Dual-path CNN and AutoEncoder framework detects adversarial intent injection in AI-native 6G networks, reaching 0.97 accuracy and 0.98 F1.

The paper defines a fine-grained threat model for adversarial intent injection in AI-native 6G intent-based networking, where malicious policies are disguised within benign intent flows. It evaluates four injection strategies: stealth-mode, random distribution, increasing frequency, and decreasing frequency. A dual-path detection framework combines a CNN using TF-IDF features for supervised detection with an AutoEncoder trained only on benign data for one-class detection, reaching 0.97 accuracy and 0.98 F1-score, roughly 9% and 36% gains over the state-of-the-art baseline.

arXiv cs.CR·Sep 10, 19:23 UTC · 6d agoResearch

35

HoneyRoute: Honeypot-Model Routing for Adversarial LLM Serving

HoneyRoute detects malicious LLM serving requests and diverts them to a honeypot model, reaching F1 0.911 with 38 ms median added latency.

HoneyRoute is an inference-serving layer pairing a streaming router (a frozen 0.8B embedding backbone with per-domain MLP heads) with a dual-implementation honeypot and an analysis loop that converts trapped interactions into attacker fingerprints for router retraining. On a production trace plus a seven-domain attack corpus it matches 96% of a two-tier guard-LLM cascade's F1 at 1/385th of its latency with 0% evasion under 13 adversarial transformations. Diverting malicious traffic cuts production token consumption under GCG-suffix flooding by 97.8%, and loop training raises detection F1 to 0.933.

arXiv cs.CR·Sep 8, 06:26 UTC · 8d agoAI safety & security

38

InceptionRAG: Stealthy Poisoning Attack Against Retrieval-Augmented Generation

InceptionRAG fragments malicious payloads into dormant passages that trigger LLMs to self-deduce misinformation via multi-hop reasoning, bypassing existing RAG poisoning defenses.

Researchers introduce InceptionRAG, a stealthy corpus poisoning attack against retrieval-augmented generation that splits a malicious payload into a chain of individually harmless dormant passages. When retrieved together, the passages induce LLMs to self-deduce target misinformation through multi-hop reasoning, achieving over 80% attack success rate across three datasets and three LLMs under rigorous adversarial constraints. A zeroth-order suffix optimization (ZOSO) method automates authoritative suffix generation in black-box settings. The authors also propose HODOR, a document isolation defense that decouples adversarial logical dependencies.

arXiv cs.CR·Sep 15, 08:22 UTC · 1d agoResearch

54

Cascading Gradient Inversion via LT-Code Inspired Peeling in Federated Learning

New gradient inversion attacks tied to erasure-coding theory recover 94–100% of ImageNet batches, showing federated learning privacy leakage is underestimated.

The paper connects gradient inversion in federated learning to erasure-correcting code theory, constructing analytic attacks that exceed previously known recovery bounds. The attacks recover batches exactly, with every sample's label, from a single FedSGD round, and certify each recovery without ground-truth data. On eight image and tabular benchmarks, even a passive attacker observing an honestly trained network recovers 94–100% of ImageNet batches up to size 128, and more than 90% actively at batch sizes of several hundred. The authors conclude that federated learning's privacy leakage has been underestimated.

arXiv cs.CR·Sep 9, 03:17 UTC · 7d agoResearch

28

Shuffling is Not Enough: Breaking Permutation-Based Model Confidentiality in Hybrid FHE Inference

Attack breaks permutation-based model confidentiality in hybrid FHE inference, recovering all ResNet-20 linear layers exactly with d+1 queries per layer.

The paper shows output-permutation plus noise fails to protect model confidentiality in hybrid FHE inference: d+1 admissible queries recover an exact permutation-invariant summary of a d-input linear layer, and shuffle-model DP amplification premises cannot hold under correctness-bounded noise. The authors recovered all linear layers of a Safhire-style ResNet-20 end-to-end from TFHE transcripts with zero error, using 5,712 total queries. Exact per-layer recovery was also confirmed on pretrained ImageNet-scale CNNs and ViT-B/16. Leaked layer spectra enable model fingerprinting, lineage attribution, and improved logit-based extraction, while suppressing them destroys inference utility.

arXiv cs.CR·Sep 11, 14:37 UTC · 5d agoResearch1

45

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled pure-autoregressive testbed shows task-specific validation losses rank image tokenizers differently, with I2T loss the most consistent signal.

Researchers built a controlled pure-autoregressive testbed and tracked task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They find losses should be analyzed per task because they exhibit distinct scaling behavior and rank tokenizers differently, and that the loss-performance relationship depends on the predicted token space. I2T loss, computed over a shared text vocabulary, correlates consistently with both generation and visual understanding performance after supervised finetuning. Case studies revisit the discriminator, semantic supervision, and vocabulary size as tokenizer design axes.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:57 UTC · 8d agoAI research1

30

Algorithmic stability via ensembling

Theoretical work derives a general framework quantifying stability guarantees for averaging-based ensembles under arbitrary data perturbations via covariance operator norms.

The paper develops a framework for quantifying algorithmic stability of ensembling strategies defined via averaging, for varied types of data perturbation. The main result bounds the stability of the ensembled algorithm in terms of the norm of a covariance operator describing the ensembling process. The framework yields interpretable insights across practical perturbation examples and provides sharper guarantees than those derived from differential privacy considerations.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:39 UTC · 7d agoAI research

22

AdamX: Cosine similarity meets gradient descent

Researchers propose AdamX, a cosine-similarity-based first-order optimizer with variance rectification that matches Adam-class convergence across benchmark datasets and architectures.

The paper introduces AdamX, a first-order optimizer that uses cosine similarity as an adaptive mechanism for controlling update magnitudes, plus a variance rectification scheme for smoother optimization early in training. The method is described as scalable, model-agnostic, and straightforward to integrate into existing pipelines. Empirically, AdamX shows competitive convergence rates measured by epochs to reach performance thresholds under a fixed hyperparameter budget, with code and experiments released on GitHub.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:42 UTC · 6d agoAI research

22

Studying Image Tokenizers as Visual Languages in Unified Multimodal Models

A controlled autoregressive testbed shows validation losses must be analyzed per task, and image tokenizer choice affects joint multimodal text modeling.

Researchers built a pure-autoregressive testbed to study image tokenizers as the 'visual language' of unified multimodal models, tracking task-specific validation losses during multimodal continual pretraining across text, image, text-to-image (T2I), and image-to-text (I2T) prediction. They found that losses exhibit distinct scaling behavior per task and rank tokenizers differently, and that I2T loss over a shared text vocabulary gives a more consistent loss–performance signal than T2I loss. Better reconstruction does not necessarily yield lower task-specific losses or stronger downstream performance, and tokenizer choice can affect text modeling under joint optimization. Case studies examine the discriminator, semantic supervision, and vocabulary size design axes.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research1

40

When Does Scale-Invariant Optimization Become Unstable? An Exact Schedule Law with Weight Decay

Researchers derive an exact law linking learning-rate schedules and weight decay in normalized networks, pinpointing when scale-invariant optimization destabilizes.

The paper shows that normalization makes large parts of neural networks scale-invariant, creating a hidden feedback loop where learning-rate schedules and weight decay interact through the parameter norm to control the effective optimizer step. An exact discrete-time law with a single scalar quantity separates contraction- and expansion-dominated effective learning-rate regimes, and the balance point is intrinsically unstable, so constant learning rate with weight decay produces recurrent behavior instead of a stable equilibrium. A unified homogeneous-optimizer framework explains why adaptive methods stabilize more weakly under normalization. The law is validated with high precision on MLPs, CNNs, and GPT-2 across MNIST, CIFAR, WikiText, and OpenWebText, with code released on GitHub.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:46 UTC · 8d agoAI research

30

Understanding the Impact of Model Pruning on Long-Tail Forgetting and Explanation Reliability in Medical Imaging

Systematic study finds model pruning causes frequency-dependent long-tail forgetting in medical imaging and that gradient-informed methods best preserve explanations.

Across two long-tailed medical imaging datasets, two CNN architectures, four pruning methods, and sparsity up to 95%, the study measures predictive performance, explanation stability, and faithfulness. Rare classes degrade earlier and more severely than frequent ones, while explanation reliability depends mainly on the pruning strategy, with gradient-informed methods degrading least. Mechanistic analysis ties explanation collapse to loss of class-discriminative gradients rather than vanishing feature activations, recommending class- and explanation-aware evaluation of compression.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:44 UTC · 9d agoAI research

27

Certifying Adversarial Robustness of Quantum Classifiers under Known-Readout Query Access

Framework certifies adversarial robustness of quantum classifiers using only measurement statistics and finite-shot outcomes, demonstrated on IBM Quantum hardware.

The paper introduces a measurement-only certification framework for adversarial robustness of quantum classifiers under known-readout query access, requiring no tomography, parameters, or gradients. It returns a lower bound ruling out untargeted errors within a radius and an attack-independent upper bound witnessing an adversarial state, both estimable with finite-sample guarantees. Evaluations show the lower bound tracks exact optima on tractable instances while the upper bound stays informative when standard attacks fail. The method was validated on IBM Quantum hardware using 40 executions of two 8-qubit quantum neural networks.

arXiv cs.CR·Sep 10, 14:45 UTC · 6d agoResearch

27

When Agents See Differently: Exposing UI Desynchronization Threats in Mobile Agents

Researchers expose 'human-agent UI desynchronization' attacks where repackaged APKs invisibly mislead mobile AI agents into attacker-chosen actions.

The paper introduces human-agent UI desynchronization: agents ingest digital screenshots and accessibility metadata that reveal content human users cannot perceive due to occlusion and luminance-contrast limits. An automated framework embeds perturbations into repackaged APK clones that steer mobile agents toward attacker-designated actions without access to runtime user instructions or online adaptation. Evaluations across five mobile-agent frameworks and three backbone models on 546 tasks achieved average misleading rates of 77.9% and 66.9%. A questionnaire study with 186 participants found the visual perturbations difficult for humans to notice.

arXiv cs.CR·Sep 15, 07:02 UTC · 1d agoAI safety & security

52

Windows Malware Detector as a Compound AI System: Trade-Offs in Accuracy, Efficiency, and Adversarial Robustness

Researchers model industrial Windows malware detection as a Compound AI System, quantifying trade-offs between detection accuracy, efficiency, and adversarial robustness under tiered attacker knowledge.

Industrial Windows malware detectors combine rule-based mechanisms with ML-based static and dynamic analyses, but their architectures are rarely publicly disclosed. The authors propose a methodology balancing detection performance, computational cost, and robustness, plus system-level threat models capturing whole-pipeline evasion rather than isolated components. Experiments on real-world data show training time reductions with marginal detection loss, while more knowledgeable attackers craft increasingly effective adversarial examples. The paper derives deployment guidelines from the observed efficiency-robustness trade-off.

arXiv cs.CR·Sep 8, 08:04 UTC · 8d agoResearch

42

A Deep Generative Model for Synthesizing Labeled Wireless Signals

Researchers propose IIns-GAN, a GAN that synthesizes realistic labeled ultra-wideband wireless signals, cutting dataset costs for wireless sensing training.

The paper introduces Inter-Instance Generative Adversarial Networks (IIns-GAN), a deep generative method that synthesizes realistic wireless signals with position-related labels to avoid costly real-world measurement and labeling. Unlike environment-model-based synthesis, the generated signals adapt to different environment scenarios and support training tasks such as distance estimation and environment identification. Experiments on public Ultra-Wideband (UWB) datasets show the synthetic signals closely mirror real measurements and improve model training performance.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:44 UTC · 12d agoAI research

20

Privacy Failure in Split-LLM Training, The Returned Gradient Nullifies the Decoys

Researchers show split-LLM training leaks privacy via zero-valued gradients on decoy rows, exposing which activations are real despite passing forward-channel checks.

A systems-security case study of a two-node split-LLM training setup found that the returned output gradient from an Untrusted Cloud Node is exactly zero for decoy rows, revealing which rows are real. Across nine seeds, zero patterns identified real rows in 4,096 of 4,096 frames per run, and an attack on frame contents recovered 0.65 to 1.50 percentage points of extra tokens over a baseline. Both datasets passed forward-channel privacy and quality checks but failed once the returned gradient was included. Row-wise gradient clipping and noise closed the leak for roughly 0.01 nats of held-out cross-entropy, though five unmeasured attack classes remain.

Hugging Face daily papers·Sep 2, 20:00 UTC · 14d agoAI safety & security

58

RelightFormer: Feed-forward Generative Transformer for Multiview Object Relighting

RelightFormer is a feed-forward generative transformer for photorealistic single- and multi-view object relighting, trained on a 90K-object dataset.

Researchers introduce RelightFormer, a feed-forward generative transformer adapted from a video foundation model that performs direct image relighting without explicit intrinsic property estimation. The architecture injects target environment maps via a latent illumination module with cross-attention and uses permutation-invariant positional encodings for unordered multi-view inputs. Training relies on the newly constructed Laval Objaverse Dataset (LOD) with 90K objects and 39K unique illuminations, and the model shows state-of-the-art quality with strong zero-shot generalization across single-view, multi-view, and novel-view relighting.

Hugging Face daily papers·Sep 6, 20:00 UTC · 10d agoAI research

38

Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data

Study finds Mixture-of-Experts models overfit faster than dense Transformers under repeated training data, with degradation tied to total parameter sparsity.

Across models from 80M to 1B active parameters (8.5B total), MoE architectures degrade more rapidly than dense models when training data is repeated, with the effect increasing with sparsity as dictated by total parameters. Dense 80M models tolerate 8x repetition with minimal loss while MoEs suffer at 4x and underperform dense models beyond 32x. Masking-based regularization such as dropout mitigates overfitting, letting MoEs beat dense models even at over 64x repetition, though no method matches all-unique training data. Routing stabilizes early and expert specialization correlates with overfitting to repeated data.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:57 UTC · 6d agoAI research

36

Competence-Gated Pooling of Language Models and Priors for Event Forecasting

Paper proposes a competence gate pooling language model forecasts with external priors, improving Brier score from 0.0771 to 0.0732 across 2,357 binary questions.

The paper defines a language model's relative competence as its marginal value beyond an available external forecast, and derives conditions under Brier loss where model disagreement improves that forecast. A competence gate estimates domain-level source weights from resolved outcomes, shrinks uncertain estimates toward a global weight, and recalibrates the pooled forecast. Across 2,357 resolved binary questions and five language models, it improves the external baseline from 0.0771 to 0.0732 Brier and beats global forecast combinations, though it defers to the market on ForecastBench. Across four Qwen models, verbal confidence failed to identify when the model outperformed the external forecast, while outcome-estimated competence supported better abstention.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

38

Likelihood-free inference with nuisance parameters through normalizing flows

Researchers decompose normalizing flows to derive near-pivotal statistics for likelihood-free inference with nuisance parameters, recovering the t-test and beating Welch limits.

A new paper decomposes neural-network normalizing flows to uncover pivotal statistics in the presence of nuisance parameters using only a sample generator from the distribution of interest. The statistic is near-pivotal in the sense of minimum average KL-divergence of its p-values and can incorporate prior knowledge of group invariances such as translation and scale. Experiments show it recovers the one-sample t-test almost exactly, outperforms the Welch test on worst-case size over a constrained variance-ratio range, and delivers higher power and much faster runtime than profile likelihood-ratio techniques on small-to-moderate samples.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 17:58 UTC · 7d agoAI research

16

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

Plan injection plants benign-sounding harmful reasoning that steers LLM actors to adversarial actions while evading chain-of-thought monitors.

Researchers show that injecting harmful but benign-sounding plans into an actor model's context causes it to perform adversarial actions while its reasoning passes chain-of-thought monitors, achieving 25-33% monitor evasion rates across benchmarks and scaling to larger models like DeepSeek-R1. Actor models paraphrase injected plans as their own reasoning without attribution. Giving the monitor access to the injected plan dropped detection by up to 50% on the Bio-Math task, with extra thinking tokens spent rationalizing rather than flagging the plan.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 18:33 UTC · 2d agoAI safety & security

55

Illusion of Depth: Revealing Hidden Stereo Vision Vulnerabilities in Depth Estimation

Simple repeating patterns let attackers shift stereo-camera depth estimates by up to 20 meters, triggering emergency braking in autonomous driving frameworks at 40 km/h.

The paper reveals an intrinsic vulnerability in stereo cameras stemming from pixel sampling and calibration processes, letting attackers finely control estimated depth of real obstacles using simple repeating patterns without adversarial ML techniques. The attack was evaluated against BM and SGBM stereo matching algorithms, deep learning models PSMNet, MoCha-Stereo, and UniMatch, the stereo-LiDAR fusion model SGM-DDC, and commercial cameras ZED2 and Intel RealSense D435; on ZED2, obstacles can be displaced up to 20 meters farther or 12 meters closer. A 0.5-second attack triggered emergency braking in a popular autonomous driving framework, with feasibility confirmed at driving speeds up to 40 km/h using CARLA. State-of-the-art defenses proved ineffective, and the authors propose a similarity-score strategy to dynamically detect and suppress depth discrepancies.

arXiv cs.CR·Sep 14, 20:51 UTC · 2d agoResearch

55