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
Researchers study Bitcoin's shift from block subsidies to transaction fees as the main mining incentive. They show that low user fees cause a mining gap, a period when miners switch off rigs, which weakens security and traps the system in an egoistic dilemma. The paper derives a strategic complementarity property among users and miners and designs an incentive mechanism based on zero-determinant theory to resolve it, with theoretical analysis and numerical simulations supporting its effectiveness.
Fix: The paper proposes an incentive mechanism leveraging the zero-determinant theory to cap the mining gap and solve the egoistic dilemma.
IEEE Xplore (Security & AI Journals)This paper proposes ATRIA, an integrity auditing scheme for multi-replica data in distributed edge computing. It uses Trusted Execution Environments to generate replica tags on behalf of resource-constrained users, aiming to resist on-the-fly and collusion attacks while adding pseudonymous identities that a Key Generation Center can trace when authorized. The authors report security analysis and a performance evaluation showing lower computational overhead than related schemes.
This paper proposes EFTE, an Entropy-Weighted Fuzzy Trust Evaluation framework for Wireless Sensor Networks. EFTE scores node trust from communication efficiency, remaining energy, behavioral consistency and mobility, using entropy-based weighting, a fuzzy inference system and adaptive thresholding to isolate untrusted nodes. Simulations and a hardware prototype on low-cost IoT devices reportedly show higher malicious node detection accuracy, fewer false positives, better packet delivery ratio and longer network lifetime.
Researchers propose a decentralized, asynchronous authentication and key agreement (AKA) scheme for vehicular ad hoc networks (VANETs), built on a consortium blockchain. The scheme targets the synchronous-network assumption of prior decentralized designs, uses lightweight cryptography and a Cuckoo filter to reduce computation, communication and on-chain storage, and reports security analyses and simulations showing robust security and high performance.
Boundary-Aware Distracted Attention Network (BADANet) addresses camouflaged object detection, where objects blend into their surroundings in color and texture. The model pairs a pretrained encoder with a boundary shrinking module, an MDASPP-based boundary map, and a multibranch bidirectional fusion block feeding a series of boundary-aware distracted attention (BADA) blocks. Across four popular COD datasets, BADANet normally outperforms 18 baseline methods, which the authors attribute to the BADA mechanism.
The paper proposes ATD, a transformer architecture for image restoration that models global dependencies with linear complexity relative to image size. It uses a learnable token dictionary and token dictionary cross-attention (TDCA) to group features into categories, and reports state-of-the-art results on multiple image super-resolution benchmarks, plus ATD-U for denoising and JPEG artifact removal.
AIRPNet is a proposed network that performs image restoration inside the steganographic domain, so the secret image stays hidden throughout processing and user privacy is protected. It uses a wavelet lifting-based Adaptive Invertible Hiding module to conceal a low-quality secret image in a stego image, and an adaptive secure restoration module to handle multiple degradations, then extracts a high-quality secret image. The authors report that it outperforms existing methods on restoration accuracy, invisibility and security across various datasets and tasks.
This paper proposes two feature compression modes for cloud-edge multimodal 3D object detection built on a sparse tensor-based network at the edge device. Transmission-Friendly Feature Compression (T-FFC) sends only the last backbone layer output and reaches a compression factor of 4933 on KITTI with the VirConv-L network, at under a 3% loss in detection performance. Accuracy-Friendly Feature Compression (A-FFC) adds two feature types, compressing by about 733 with almost no performance loss.
This study proposes an outlier-aware contrastive learning method for high-fidelity embeddings. It addresses sampling bias, where similar samples from different classes create false negatives, by using OOD detection to mask them in a debiased contrastive loss. Because prior knowledge of OOD samples is usually unavailable, the authors generate synthesized OOD samples near the boundary between in-distribution (ID) and OOD data, learning an OOD detector from these samples and ID data. The source reports experiments illustrating the merit of the approach.
This paper proposes a distributed multi-attribute anonymous certificate management scheme for uncrewed aerial vehicles (UAVs), which need certificates carrying multiple attributes as proof of mission authorization. It combines redactable signatures and dynamic accumulators to allow selective disclosure and fine-grained revocation of individual attributes within one certificate. It also uses a distributed key-generation mechanism so certificates can be issued in a decentralized way, avoiding the need to apply to separate centralized authorities for cross-regional access.
The paper introduces PSDO (Privacy-Sensitive Distributed Optimization), a decentralized energy management framework for prosumers that combines decentralized optimization with differential privacy. Tested on the IEEE 33-bus radial distribution system, PSDO converges to the optimal solution while upholding differential privacy, and it outperforms the baseline DP-ADMM with superior privacy protection and minimal utility loss.
Human annotators often disagree, which introduces label noise into classification datasets, and the true noise distribution is usually unknown. This research shows how interannotator agreement statistics can estimate that noise distribution and uses those estimates to train models on noisy data. It also derives generalization bounds within the empirical risk minimization framework that depend on the estimated noise characteristics.
OwnerHunter is a multilingual framework that uses large language models to identify the owner of a website, treating it as a document-level information extraction task rather than sentence-by-sentence named entity recognition. The authors evaluate it on the refined English ONER dataset and a new Chinese dataset, WOI-cn, containing 16,036 real websites. It reports F1 scores of 0.9505 on ONER and 0.9621 on WOI-cn, which it describes as new state-of-the-art results.
Researchers propose a hybrid detector that combines an enhanced Character-Level CNN (CLCNN) with a reinforcement learning module to catch adversarial SQL injection payloads that evade Web Application Firewalls. The CLCNN extracts multi-scale features, and the RL module de-obfuscates payloads where the CLCNN's confidence is low. The paper reports that the CLCNN resists tools such as AdvSQLi, RAT, and GPTFuzzer without prior knowledge, and that RL-based adversarial training raises Recall with minimal loss on standard SQLi datasets.
CipheRAG is a privacy-preserving retrieval-augmented generation framework for large language models that aims to balance knowledge confidentiality with retrieval efficiency. It combines a searchable inner product functional encryption mechanism, enhanced with asymmetric locality-sensitive hashing, with a decryption-enabled attention mechanism that feeds decrypted knowledge into the generation process. The authors report up to 35x faster generation and 15x faster QKV computation than FHE- and OT-based baselines.
This paper argues that the goal of safe exploration in reinforcement learning is to reach an equilibrium between the feasible zone and the environment model, since each improves the other. It proposes safe equilibrium exploration (SEE), which alternates between finding the maximum feasible zone and the least uncertain model, and proves that the model is monotonically refined and the feasible zones monotonically expand toward equilibrium. On classic control tasks, the authors report that SEE expands feasible zones with zero constraint violation and reaches equilibrium within a few iterations.
The article examines how threat actors may use on-device large language models inside living-off-the-land (LOTL) attack pipelines, where attackers rely on legitimate tools and processes already present on a system to avoid detection. It also addresses how to mitigate this threat.
The paper presents AdvDiffusion, a method for generating adversarial face patches that can be pasted onto a face to cause face recognition models to make false identity judgments. It first selects a patch region using facial gradient maps, then adds noise to an image and denoises it with a pre-trained diffusion model, guided by an adversarial loss. Experiments report effective attacks on black-box face recognition models in both digital and physical settings, with better robustness to complex physical environments than some state-of-the-art methods.
The authors examine how adversarial perturbations shift text-guided attention in CLIP, a pre-trained vision-language model. They propose TGA-ZSR, which uses a Local Attention Refinement Module and a Global Attention Constraint Module, and Comp-TGA, which adds complementary class-prompt and non-class-prompt foreground attention. Across 16 datasets, TGA-ZSR and Comp-TGA improve zero-shot robust accuracy over the prior state of the art by 9.58% and 11.95%, respectively.
The paper proposes DSW, a dataset-specific watermarking method for detecting excessive use of protected datasets when text-to-image diffusion models are fine-tuned with LoRA. An encoder embeds a watermark image representing the dataset owner's identity into the dataset samples, and a decoder trained with a bi-level optimization strategy extracts that watermark from images generated by the fine-tuned model. The authors report experiments validating the availability, effectiveness, and stealthiness of DSW.