Research
Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
Academic papers, new techniques, benchmarks, and theoretical findings in AI/LLM security.
2,040 items
This article compares Service Mesh and Event Mesh approaches for microservice-based applications deployed across the cloud continuum. It argues that Event Meshes, which promote event-based asynchronous interactions among services and data, improve decoupling and optimize performance in traffic management, observability, and security.
This article examines how AI techniques, including advanced machine learning, computer vision, and generative models, can be combined with sensing technologies, remote monitoring, and geospatial analytics to find untapped natural resources. The authors argue that AI-driven exploration reduces human risk and ecological impact compared with conventional methods, which are costly, slow, and environmentally disruptive. The source text reports no quantitative results.
The paper proposes DQEN, a dual query enhancement network for DETR-based human-object interaction (HOI) detection. It replaces randomly initialized object and interaction queries with object-aware encoder features and semantic features from a contrastive language-image pretraining model, plus an auxiliary prediction unit. The method reports competitive performance on the HICO-Det and V-COCO datasets.
MSFFNet is a lightweight vehicle reidentification network that fuses multistage CNN features through a fusion head and a graph convolution-based fusion block. A metadata auxiliary module injects camera IDs and viewpoints to suppress irrelevant information. The authors report state-of-the-art results on three vehicle ReID benchmarks at 1.74 ms per image.
The authors propose FLAT, a federated learning scheme that uses decentralized secure aggregation and a FedAvg algorithm built on secure multi-party computation, with a new dropout handling mechanism. They show that maximum communication complexity at individual nodes drops from O(pn) to O(p√n). Experiments report execution speedups of 2.7x-4.3x in a LAN and 3.5x-3.85x in a WAN.
STMWF is a multi-tab website fingerprinting attack against Tor-encrypted traffic that jointly models Inter Arrival Time (IAT) with the direction sequence rather than analysing them separately. It uses an SE-attention feature extractor, correlation self-attention and correlation denoising to build per-website fingerprints. Compared with the state-of-the-art multi-tab attack ARES, it improves average F1-score by about 14.87% under static conditions and 34.81% under dynamic conditions, and it is more robust against WF defenses.
TrapFlow is a website fingerprinting defense based on backdoor learning. It injects crafted trigger sequences into targeted website traffic so that an attacker's model learns incorrect associations, whether it retrains on the noisy data or trains only on clean data. In experiments, it cut the accuracy of the RF attack from 99% to 6% with 74% data overhead, and the authors also validated it in a real Tor network environment.
Researchers ran a qualitative multiple-case study of 12 Chinese SMEs across four platform ecosystems, drawing on 54 semi-structured interviews and archival data. They identified four resource dependence mechanisms (tangible, intangible, human, and data) that shape platform strategy. The main finding is that concerns over dependence, rather than resource scarcity, primarily drive platform choices.
Researchers investigate why private individuals in the United States do or do not adopt personal cyber insurance, which covers residual cyber risks. Using survey data from 301 individuals and Fuzzy-Set Qualitative Comparative Analysis (fsQCA), they test 12 a priori propositions drawn from Protection Motivation Theory (PMT) and Theory of Planned Behavior (TPB). The study finds asymmetric explanatory patterns that go beyond symmetric, variance-based models of insurance adoption.
FedPerX is a federated transformer framework for multilingual sentiment analysis that combines residual adapter-based personalization with adaptive multi-granular differential privacy. It keeps a frozen XLM-R backbone and trains lightweight client-specific adapters, with noise injected at the feature and adapter levels. On the MARC and TSMD benchmarks, it reports up to +4.3% macro-F1 gains over seven baselines and a 70% reduction in communication overhead.
The paper proposes RCMCL, a Robust Trusted Conflictive Multiview Collaborative Contrastive Learning method, to address multiview learning models that ignore decision uncertainty and degrade on conflictive instances caused by misaligned multiview data. RCMCL combines evidential deep neural network views with dissonance-based evidence contrastive learning and collaborative learning of consistent and complementary evidence. The authors report that experiments on eight benchmark datasets show RCMCL outperforms state-of-the-art methods.
The paper presents ADVersa, a framework that infers plausible visual and textual explanations for absent near-crash scenes in driving accident videos. It introduces MM-AU, a dataset of 11,727 in-the-wild accident videos with aligned text descriptions, 2.23 million object boxes, and 58,650 video-based accident cause text pairs. The authors report that ADVersa outperforms state-of-the-art methods on tasks including near-crash frame recovery, frame prediction, cause reasoning, and video synthesis.
The authors propose EasyHard-FSD, a fingerprint spoof detection method that addresses bias toward easy-to-recognize samples caused by nonuniform spoof representations from different artificial materials. It uses a teacher/student self-training setup in which a teacher model, trained on all original samples, identifies hard samples, and a hard mining module (HMM) masks those samples for the student model. The method reportedly outperforms state-of-the-art approaches on the LiveDet2013 and LiveDet2015 datasets.
This paper proposes SDkA, an anonymity model that combines synthetic data with k-anonymity for data sharing. Synthetic data from a conditional generative adversarial network, together with an automated quality-fixing method, addresses size, quantity and diversity issues. Selective generalization is added to the k-anonymity model to avoid unnecessary generalization, and the authors report that SDkA defends against diverse attacks and improves utility over standard k-anonymity.
The article argues that artificial intelligence and other emerging technologies now play an overwhelming role in modern business, making most companies technology companies to a significant degree. It holds that executives need a conceptual understanding of these technologies, especially AI, to pursue digital transformation effectively and to anticipate their impact on their companies.
This survey synthesizes jailbreak research on large language models from 2023 to 2025, covering attack methods, defense strategies and evaluation frameworks. It groups jailbreak techniques into five categories: prompt-based injections, role-play conditioning, multiturn dialogue, multilingual or multimodal exploits, and optimization-driven pipelines. It also reviews defenses such as SFT, RLHF, adversarial fine-tuning, and output or pipeline filtering, and analyzes benchmarks including PromptBench and JailbreakBench.
Researchers study distributed black-box optimization, where client objective functions can only be queried, and address slow convergence caused by data heterogeneity and high gradient-estimate variance. They propose DZO-SNGM, a distributed zeroth-order stochastic normalized gradient descent method with momentum, which runs multiple local zeroth-order updates per client, aggregates models globally, and applies one stabilizing momentum update on the server. Experiments on image classification and distributed black-box adversarial attacks report faster convergence and better final performance than state-of-the-art methods.
The paper proposes the Cost-efficient LandscapE Flattening (CLEF) attack, which improves targeted adversarial transferability by optimizing perturbations at both local maxima and local minima of the input loss landscape. CLEF reuses gradients from the previous attack step to reach maxima and uses a probabilistic model, pre-trained on dozens of images from other domains, to sample perturbations that reach minima. Experiments report that this flattens the loss landscape of crafted adversarial examples and improves their transferability.
The authors present a detector for AI-generated images that combines a sliding-window attention mechanism, which models relationships between neighboring elements within local regions, with a dual frequency branch that uses four DWT subbands and the phase part of FFT. Evaluated on datasets covering images from 65 generative models, including GAN and diffusion-based ones, the method improves detection accuracy by 2.13% over state-of-the-art approaches.
AdvScan is a black-box method for detecting adversarial examples on TinyML models at runtime by analyzing power consumption. It builds a baseline of power signatures from benign inputs and applies a one-sample t-test to flag inputs whose signatures deviate significantly. Evaluated on three MLPerf Tiny models across two STM32 microcontrollers and 318,400 test inputs, it detected 99.984% of adversarial examples, with 40 false negatives and no false positives.