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 is the release candidate (RC) 2025-11-25 revision of the Model Context Protocol, with the specification available in draft form and a changelog linked for details. The source states the spec is not final and changes may occur before the final release, and that SDKs will adopt it at their own pace.
Researchers propose DIC-GAN, a generative adversarial network that reconstructs radar reflectivity mosaic images from multichannel geostationary satellite observations. The generator uses nine dynamic identity convolution modules and a mixed loss combining BCE, MSE and weighted threshold terms to better capture strong reflectivity regions. The authors report that DIC-GAN outperforms comparison methods and generalizes to areas without radar coverage.
HPE-Li++ is a Wi-Fi and camera-based approach to 3-D human pose estimation designed to run on resource-constrained devices. It combines a multibranch CNN with selective kernel attention, a Transformer module and a stacked autoencoder denoiser. On the MM-Fi and WiPose datasets it reports 85.58% and 94.27% PCK50, outperforming state-of-the-art methods, and keeps 80% PCK50 under AWGN noise with an error variance of 0.5.
Researchers propose FGE-GAN, a fuzzy graph evolutionary generative adversarial network, to predict Alzheimer's disease risk and extract pathogeny. The method models disease progression as fuzzy entropy spreading through fuzzy graphs, using fuzzy graph convolution layers in the generator. Experiments on multiple brain disease datasets reportedly show it outperforms state-of-the-art methods in risk prediction, and the extracted multiomics pathogenies are said to offer insights for early intervention.
Researchers propose TEDD, an ensemble model for link prediction in dynamic knowledge graphs with drifting entities, aimed at sparse knowledge graph completion. The model combines a double scoring mechanism that uses both text semantics and triple structure, and a dual-BERT design with an auxiliary candidate entity set to avoid redundant embedding calculations. The authors report that TEDD outperforms state-of-the-art methods on benchmark datasets.
Researchers propose ABAE-RTN, a deep learning framework for physical layer security in wireless networks. It combines adaptive beamforming, radio transformer networks and autoencoders, and injects controlled artificial noise to disrupt eavesdroppers. On the DeepMIMO dataset, across static (O1) and UAV-assisted dynamic (O1-Drone) environments, it reportedly improves secrecy rate, bit error rate and outage probability over LSTM and GNN baselines.
MITRE released ATLAS Data v5.1.0, which contains 1 matrix, 16 tactics, 84 techniques, 56 sub-techniques, 32 mitigations, and 42 case studies. The release adds the Lateral Movement tactic and new techniques such as Manipulate User LLM Chat History and Delay Execution of LLM Instructions. It also adds six mitigations covering AI agent permissions, tool invocation, human-in-the-loop review and memory hardening.
Researchers studied memristive Hopfield bi-neurons under electromagnetic radiation and stimulating currents, finding stable, unstable, limit-cycle and multi-scroll chaotic states. A digital hardware device reproduced these patterns. The chaotic sequences were then used with compressive sensing to compress and encrypt biomedical images, which the authors report is robust against various attacks with a high key space.
Researchers present FedFFTNet, a lightweight dual-branch deep learning architecture that identifies source camera models from pattern noise using federated learning, so training data stays private. A Laplacian-based patch selection strategy focuses on sharp image regions to improve feature extraction. On four benchmark datasets (Forchheim, Dresden, Vision, Socrates) under the federated setup, the system reports image-level accuracy of 99.10%, 99.78%, 98.64% and 97.38%.
Standard integral neural networks are hard to deploy on resource-constrained mobile devices because their training-time and inference-time structures are coupled. The authors propose a continuous reparameterization strategy that converts multiple training-time integration layers into a feed-forward structure at inference, and build lightweight reparameterizable INNs (RINNs) on a MetaFormer-style design. On ImageNet, RINNs reach over 79.1 top-1 accuracy at 0.87 ms latency on mobile devices, and keep their performance at up to 50% structural pruning without fine-tuning.
Co-AttenDWG is a multimodal learning approach that projects text and visual features into a shared embedding space and uses coattention with dimension-wise gating and expert fusion to model cross-modal interactions. Its dual-path encoders and cross-attention layer refine and align modality-specific representations before an expert fusion module combines them. The authors report state-of-the-art performance and better cross-modal alignment on the MIMIC and SemEval Memotion 1.0 datasets.
This paper introduces a federated learning approach for estimating the structure of a dynamic Bayesian network from time series data split horizontally across parties. It extends the method to heterogeneous data using a proximal operator as a regularization term in a personalized framework, and proves convergence to a stationary point. Experiments on synthetic and real-world datasets show it outperforms state-of-the-art techniques, especially with many clients and small per-client samples.
MITRE released ATLAS Data v5.0.0, which adds a "Technique Maturity" field to the distributed ATLAS.yaml file. The field grades evidence behind each technique as Feasible, Demonstrated, or Realized. The release also adds several new AI agent techniques, including AI Agent Context Poisoning, RAG Credential Harvesting, and Exfiltration via AI Agent Tool Invocation, plus one new case study.
This article presents a maritime operational technology shipboard testbed for cybersecurity research and development. The testbed offers a realistic environment for safely benchmarking cyberattacks and defenses.
This research paper argues that the continuity of a classifier, rather than robustness and accuracy themselves, is what conflicts with them in many scenarios. The authors introduce a framework for studying harmonic and holomorphic hypotheses in learning theory terms and report empirical evidence that continuous hypotheses underperform discontinuous ones on some common machine learning tasks. They explain adversarial examples as a conflict between continuity and uniform convergence to a discontinuous function.
Researchers present FUBA, a backdoor attack on federated learning models that exploits federated unlearning. Adversarial clients inject a backdoor during training, then one client submits an unlearning request to remove its contribution, which activates the backdoor in the model. The authors report that FUBA reliably backdoors the model after unlearning regardless of the unlearning method used, and that it remains stealthy against existing backdoor defenses.
Authors Alex Irpan, Alex Turner, Mark Kurzeja, David Elson and Rohin Shah propose consistency training, a self-supervised method that teaches a model to ignore irrelevant cues such as user biases or jailbreak wrappers. They introduce Activation Consistency Training (ACT), which optimizes internal activations, and compare it with Bias-augmented Consistency Training (BCT) and baselines on Gemma 2, Gemma 3 and Gemini 2.5 Flash, reporting that ACT and BCT beat the baselines and improve robustness. The approach avoids the static datasets that SFT relies on, which can go stale in guidelines or capability.
MaxDiv is a zero-shot machine unlearning method that erases knowledge from a model without access to its original training data. It generates erasing samples by negating the distributions of the data to be forgotten, and adds knowledge distillation to prevent catastrophic forgetting of retained information. Evaluations on MNIST, SVHN, CIFAR-10, CIFAR-100 and TinyImageNet reportedly show superior performance over current state-of-the-art methods.
This systematic literature review applies a SWOT framework (strengths, weaknesses, opportunities, threats) to prompt engineering techniques for large language models (LLMs). It covers how designed inputs such as questions or instructions steer model responses, with applications discussed in healthcare, education and customer service. The authors highlight challenges including maintaining accuracy and efficiency.