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.
2,033 items
This paper proposes RD-PCN, a role-differentiated payment channel network for blockchain cryptocurrencies, after a measurement study showed node heterogeneity in real-world PCNs. It introduces multi-party payment channels, including a PCN-compatible design called PC-MPC, to raise channel fund utilization, and outsources routing to resource-rich nodes using trusted execution environments. Experiments report at least an 18.9% higher payment success ratio and 60% higher fund utilization than representative schemes.
Researchers study the weight stable set attack problem with budget constraint, where an attacker removes network nodes within a removal-cost threshold to minimize the maximum weight stable set in the remaining graph. They propose a 2α-approximation algorithm for graphs without odd cycles, where α is the approximation ratio of the minimum knapsack algorithm, and extend it to general graphs alongside a genetic algorithm. Experiments on artificial and real-world networks show a performance ratio below 1.21 for graphs without odd cycles.
The paper introduces Active Weighted Signature (AWS), a scheme for Proof of Stake blockchains that adjusts participant weights and thresholds without trusted dealers or re-setup. It instantiates AWS as ADAPT by applying Generalized Lagrange Interpolation (GLI) to the Schnorr-based threshold signature FROST. Implementation results show key generation efficiency comparable to FROST, weight and threshold adjustments completing in 4.1-22.3% of re-setup time, and sub-linear scaling for uneven weights: a 49× weight difference requires only 3.29× computation versus 49× under virtualization.
This paper introduces ADASM, an approach that uses adversarial domain alignment to improve how stereo-matching networks generalize from synthetic training data to real-world data. It combines adversarial training with Maximum Mean Discrepancy to align the original domain with a fictitious noisy domain. The authors report superior disparity estimation on KITTI, Middlebury, and DrivingStereo, with performance competitive with or better than fine-tuning and no fine-tuning needed.
An ethnographic study conducted in Italian schools examines how the General Data Protection Regulation applies to primary and secondary institutions. The authors discuss the tradeoff between top-down compliance and risk-based operational resilience.
The source argues that students increasingly skip books in favor of free videos and AI tools, leaving security books by experts unread. It presents books as a continuing source of security knowledge.
Researchers propose a blockchain voting protocol that supports quadratic voting with liquid democracy, letting voters delegate to representatives while splitting their voting power across selected projects. They model the protocol in the Universal Composability framework, prove UC-security under the Decisional Diffie–Hellman assumption, and report a prototype where a setting with 64 voters, 128 delegates and 128 projects produces about 1.4 MB of traffic per voted project.
Clara is a cross-modal learning framework for software vulnerability detection that combines code sequences and code graphs, using large language models and graph neural networks. It adds a prompt-guided information interaction module for local fusion and a Cross-attention Adaptive Fusion module for global fusion. On two benchmark datasets it improves F1 score by 21.37% and 11.86% over state-of-the-art vulnerability detection methods.
The paper introduces the Cryptographic Resource Enabled Framework (CREF) for private-keyless data custody, where one-time attribute credentials issued by multiple authorities replace user private keys. It builds a Credentials-Driven Cryptosystem (CDC) over ideal lattices, using a preimage Gaussian sampler for short attribute credentials and a small policy matrix to convert monotone policies into cryptopolicies. The authors claim existential unforgeability of credentials and cryptopolicies and lower storage and computational costs than existing schemes, without reencrypting data when policies change.
The paper examines wireless challenge-response Physical Layer Authentication (PLA) built on OFDM, which assumes a sufficiently random fading channel. It shows that OFDM subchannel responses can be correlated in practical settings, and proposes the Maximum Differential Likelihood Generator (MDLG), an adversary model that exploits this weak correlation to attack PLA. Real-world experiments validate the attack and a randomness-testing guideline for deciding when PLA is safe to use.
The authors propose a searchable encryption scheme that provides system-wide security for multi-keyword retrieval, using volume bucketization to protect the volume of retrieved documents and dynamic bitmap and adaptive volume hiding techniques for Boolean queries. They report that on large-scale datasets their method outperforms existing solutions by at least 100× in write-back and update efficiency, cuts client-side cache overhead by 78%, and processes a Boolean query with 8 keywords in approximately 454.78 ms.
This study examines whether speaker-adversarial perturbations, which mislead speaker recognition models, can be removed from speech entirely. The authors test three scenarios of attacker knowledge (ignorant, semi-informed, well-informed) across optimization-based and feedforward perturbation generators on the LibriSpeech dataset. They find perturbations are not eliminated in the ignorant or semi-informed cases, while feedforward perturbations are considerably reduced in the semi-informed case and nearly eliminated in the well-informed case, allowing restoration of the original speech.
The source argues that evasion attacks on machine learning systems deserve more concern than their usual treatment suggests. Canonical demonstrations have stayed largely academic for over a decade, and their intellectual appeal has led many to dismiss evasion as having little operational consequence.
This article describes a hybrid triage pipeline that combines static application security testing with large language models. The pipeline provides intelligent triage and automated exploit generation. According to the source, it reduces false positives by 91% in real-world deployments.
SAFER is a method for removing target concepts from text-to-image diffusion models so they do not reappear in subtle forms. It identifies a concept-specific subspace in the text embedding space and projects prompt embeddings onto its complementary subspace, using textual inversion from a reference image and a subspace expansion strategy. The authors report that it consistently erases unwanted concepts while preserving generation quality.
Researchers propose a proactive Deepfake detection framework that embeds a watermark into the Quaternion Radial Harmonic Fourier Moments (QRHFMs) domain of face images. A Local-Enhanced State Space (LSS) block extracts features, and a dual-branch architecture pairs a watermark extractor with a forgery discriminator via knowledge distillation. On benchmark datasets the method reports 92.58% average accuracy and 95.96% AUC, with robustness to image quality degradation and to Deepfake techniques unseen during training.
Fix: The paper proposes a measurable guideline based on randomness testing to decide when PLA can be used under practical wireless channel conditions.
IEEE Xplore (Security & AI Journals)R$^{2}$2BD is a fake image detection framework that uses a unified reconstruction model, G-LDM, which simulates the generation behaviors of VAEs, GANs and diffusion models. Its residual bias calculation module distinguishes real from fake images in a single inference step, where existing reconstruction-based methods typically require 20$+$+ steps. Across 10 public datasets it is reported to be over 22× faster while achieving superior accuracy, and in cross-dataset evaluations it outperforms state-of-the-art methods by an average of 13.87%.
This paper introduces an Embedding Reconstruction Attack (ERA) framework for DNN-based face and speaker recognition systems, covering a four-tier taxonomy of adversary access from score-only and decision-only interfaces down to no-query/no-feedback scenarios. Using Stable Diffusion for faces and flow-matching transformers for voices, the reconstructors achieve 93.27%, 82.60%, and 62.29% success for Tiers 1–3 under identical query budgets, outperforming existing black-box attacks. The authors prove Tier 4 is infeasible under their threat model, and find compressed (pruned, quantized, distilled) models remain at high risk.