Research

Researchers study why deep image forgery detectors fail to generalize to unseen GANs and to noisy samples. They trace both problems to a frequency bias in DNN detectors and propose a two-step frequency alignment method that removes the frequency discrepancy between real and fake images. The method works as a black-box attack on detectors or as a universal defense, tested across twelve detectors, eight forgery models and five metrics.

The authors propose a privacy-preserving proxy bilateral access control scheme for secure data forwarding. Proxy re-encryption alone lets only the data owner set who can decrypt, and forwarding through multiple nodes exposes data to impersonation and forgery. The scheme lets both sender and receiver state access policies, and the authors prove message confidentiality and authenticity under standard assumptions in the random oracle model.

The Model Context Protocol has published its stable 2025-11-25 revision. The specification is available on the official Model Context Protocol website, and a changelog describes the changes in detail.

GRACE-FL is a federated learning framework that factors device energy capacity into training, adjusting each client's learning rate, local epochs and gradient quantization bit-width to its available energy. It also uses an energy-weighted aggregation strategy meant to keep contributions fair. Experiments report 20–30% energy savings and up to 28% lower communication costs with competitive accuracy under IID and non-IID settings.

Offline reinforcement learning agents trained on fixed datasets often perform suboptimally, so they need further finetuning through online interaction. This paper proposes SUNG, a simple unified uncertainty-guided framework for offline-to-online RL. SUNG estimates uncertainty with a VAE-based state-action visitation density estimator, uses optimistic exploration, and adaptively applies conservative offline objectives to high-uncertainty samples and standard online objectives to low-uncertainty ones. It reports state-of-the-art online finetuning performance across D4RL environments and datasets.

This survey reviews deep model fusion, which merges parameters or predictions from multiple deep learning models, including large language models and foundation models, to offset individual model biases and errors. It groups existing methods into four categories: weight averaging, mode connectivity, alignment, and ensemble learning. The authors also analyze open challenges, such as high computational cost and interference between heterogeneous models, and outline future research directions.

Researchers introduce OR-SLZNet, a deep learning model based on ordinal regression that predicts safe landing zones from UAV imagery. It combines photometric cues such as color and texture with geometric cues such as flatness, slope, and depth to assign each pixel an ordinal safety score. The authors report real-time inference of about 0.02 s per frame and strong generalization across five diverse datasets.

The authors present nonlinear formulations for Shape-from-Template (SfT) and Non-Rigid Structure-from-Motion (NRSfM) that model isometric, conformal and equiareal deformations. Unlike prior methods, which rely on relaxations such as inextensibility or on optic flow around correspondences, the approach needs only point correspondences and resolves ambiguities with maximal depth and maximal isometry heuristics. Semi-Definite Programming (SDP) solves all formulations, and an adapted opposite-depth parameterisation shows a smaller relaxation gap. Experiments on real-world benchmark datasets report superior accuracy over existing methods.

Researchers propose the Grounded Cognition Method (GCM) for unbiased scene graph generation, aiming to fix the long-tailed bias that makes existing scene graph generation methods favor frequent predicates over semantically accurate ones. GCM models simulation, bodily states and situated action, using an Out Domain Knowledge Injection module, a Semantic Group Aware Synthesizer, and a Shapley Enhanced Multimodal Counterfactual module. Experiments on Visual Genome, GQA and Open Images V6 reportedly show GCM outperforming state-of-the-art methods and achieving a better trade-off.

The paper reconsiders rotation-invariant shape recognition from the perspective of contour points rather than pixels, arguing that pixel-based methods carry redundant information and leak critical geometric information early. It proposes an anti-noise rotation-invariant convolution module based on contour geometric awareness, which divides the shape contour into local geometric regions (LGA) for finer-grained rotation-invariant coding. A deep network of five cascaded modules shows excellent performance on classification and retrieval experiments and robustness to contour noise and rotation centers.

TGDIP is a two-stage graph information maximization model for drug–drug interaction prediction. It targets two problems in GNN-based methods: homogenized drug features and redundant drug-pair features. Experiments on three public datasets are reported as validating its effectiveness.

The authors introduce FedDPO, a federated learning approach for medical image classification that uses reinforcement learning to set an adaptive regularization term for each client's objective function. The method accounts for each client's data distribution, system specifications and performance feedback, and adds local batch normalization to reduce the impact of heterogeneous data. Experiments on two medical and one nonmedical image classification tasks show consistent gains over several baselines.

This paper proposes a continual learning framework for image de-raining that lets networks accumulate knowledge from a growing pool of datasets. It uses GANs to capture features of new data, trains the de-raining network on existing and GAN-synthesized data with knowledge distillation, and reports that it works across six datasets and beats state-of-the-art methods on new real-world scenes when tested on three benchmark networks.