Adversarial machine learning

The paper proposes a detector that identifies adversarial images for frozen vision-language models by profiling how an image responds to a set of general semantic prompts. The profile combines category-level statistics, prompt relationships, deviations from clean reference distributions, and stability under weak image transformations, and a lightweight classifier labels each input while the VLM stays fixed. Evaluated across multiple datasets, CLIP-style backbones and several attack families, the detector discriminates strongly in attack-specific settings and retains substantial performance on unseen attacks.

GraphRectify is a graph-based framework that transfers adversarial image detectors from one classifier backbone to another. It learns a structured representation of intermediate classifier features and adapts features from a new backbone to the detector trained on the original model. Across the evaluation matrix, it achieves higher aggregate ROC-AUC than training a detector from scratch on the new backbone, with the largest gains between different backbone families when sufficient data are available.

Researchers introduce STCA (Spatial Temporal Coherence Adversarial Attack), a black-box method that perturbs driving video to fool Vision Language Models. The attack selects semantically important frames with caption guidance, applies a spatial perturbation that preserves high SSIM, then disrupts cross-frame temporal coherence with a motion-guided mask. Tested on BDD100K and nuScenes against Video LLaVA-7B, Qwen2.5-VL-7B and Dolphin, the spatial attack reaches a high ASR while keeping SSIM high, showing these models remain highly susceptible.

This paper presents a framework for evaluating adversarial robustness and attack transferability in image-based deep learning models for malware detection and classification. The authors apply image-domain attacks from FGSM to AutoAttack and test whether binary-domain manipulations remain effective after conversion to an image representation. They report average attack success rates of 64.6% for FGSM and 98.8% for AutoAttack, and accuracy drops of up to 42% from transferred binary-domain manipulations.

Researchers study how to make adversarial images transfer from open-source surrogate multimodal LLMs to closed-source MLLMs in black-box settings. They propose IAU-FOA, which aligns adversarial and target images at both global and patch-cluster levels using confidence-adaptive unbalanced optimal transport, plus visual-invariance augmentation that simulates exposure, contrast, illumination and color-temperature changes. The authors report that it consistently outperforms state-of-the-art transferable attack methods across open-source and closed-source MLLMs.

Researchers evaluated a three-layer ensemble defense for deep learning credit risk models, combining MLP, ResNet-1D and TabTransformer architectures with PGD adversarial training and SHAP-based routing. On German Credit and Lending Club under FGSM, PGD and CW attacks across five seeds, the defended AUCs reached 0.758 and 0.723, and the default-class attack success rate fell from 0.52 to 0.19 and from 0.55 to 0.21. SHAP attribution consistency also improved, with Spearman correlation rising from 0.42 to 0.87 between clean and defended explanations.

Researchers analyze Vision-Language-Action (VLA) models with a sparse autoencoder (SAE) and find an internal feature whose activation strongly correlates with adversarial patches. They suppress this feature at inference time only when a linear probe detects an attack, which improves robustness without fine-tuning the VLA. On LIBERO-10, conditional intervention raises success rate under intermittent attacks, while continuous intervention substantially degrades policy performance.

Researchers report that a targeted adversarial perturbation can drive a vision-language model's teacher-forced training loss for a fixed target caption near zero, while free generation on Qwen2.5-VL-7B-Instruct still yields the correct description. Using 200 held-out COCO images and a two-stage PGD attack, they localise the gap to one autoregressive step and across the 28 LLM decoder layers, finding that the language decoder, not the visual encoder, differentially arbitrates which images are corrupted (linear probe AUC=0.858, flagged for circularity). The authors argue that adversarial robustness in autoregressive VLMs depends mainly on the language decoder's prior.

The paper asks whether a network intrusion detection system's model can be extracted using only hard-label black-box queries. The source text provided gives only the title, publication date (December 2026), journal (Journal of Information Security and Applications, Volume 103), and authors (Donguk Min, Seungsoo Nam, Daeseon Choi), with no method description or findings.

Researchers propose a "carrier", a secondary visual element that gives unrestricted adversarial attacks an auxiliary region to work in, so the primary object (the subject) is distorted less. Under global classifier guidance, the carrier absorbs a larger share of normalized attack updates and improves cross-model transferability. Targeted attacks keep the personalized subject as the main content perceived by humans while misleading the classifier.

Researchers present AdvPCS, a universal cross-prompt adversarial attack on Promptable Concept Segmentation in SAM3, which extends SAM-series models to concept-level prediction. The method combines min-max prompt optimization with perception deception and temporal memory misalignment attacks. A single universal adversarial perturbation (UAP) generalizes across frames from different videos and reduces the average mIoU of PCS models on the SA-CO dataset to below 5% under text prompts.

Researchers propose a Universal Adversarial Object, a sphere with an optimized surface texture, that degrades the task success of Vision-Language-Action (VLA) models when placed in a robot's field of view. Their multi-level attack framework disrupts trajectory planning, task execution and action control, and was validated in simulated and real-world robotic settings. The object reduces average task success rates by 31.2% to 39.9% for two VLA models, Pi0 and RDT, with success rates dropping to near zero in complex scenarios.

Researchers present the first systematic study of backdoor attacks against the interactivity of interactive video generation (IVG) models. Their method, BadAction, implants predefined motion patterns into action sequences so that triggered models produce frozen future frames that ignore later user actions, while benign inputs behave normally. The attack reaches average success rates of 91.0% with action-only triggers and 80.4% with multimodal triggers, and it bypasses existing backdoor detection methods.

Liu, Wen and Yu published an article in the Journal of Information Security and Applications, Volume 103, dated December 2026. The source text provided contains only the publication metadata and author list, with no abstract, method or findings.

This paper shows that an attacker can manipulate LLM behavior by poisoning the demonstration context used in in-context learning, without fine-tuning the model. The authors present ICLAttack, a backdoor method that poisons demonstration examples or demonstration prompts, reporting a 95.0% average attack success rate on OPT models across three datasets. They also propose ICLDefense, which uses a lightweight auxiliary model and an ensemble-based strategy to refine LLM outputs and reduce attack success.

Researchers propose a spectral detection method for finding adversarial attacks embedded in super-resolution model weights, a preprocessing component of imaging pipelines. The method trains an XGBoost detector on the radially-averaged power spectral density and benchmarks it against magnitude- and phase-based Fourier spectrum detectors across training and cross-architecture scenarios. The proposed detector outperforms the comparison detectors in most scenarios, and high-frequency features prove most informative for detecting AdvSR attacks.