arXiv subjects

Song Xiao

Publications and source records attributed to Song Xiao.

10 recordsLinked to original sources

arXiv2026

CMFDFormer: Transformer-based Copy-Move Forgery Detection with Continual Learning

Copy-move forgery detection aims at detecting duplicated regions in a suspected forged image, and deep learning based copy-move forgery detection methods are in the ascendant. These deep learning based methods heavily rely on synthetic training data, and the performance will degrade when facing new tasks. In this paper, we propose a Transformer-style copy-move forgery detection network named as CMFDFormer, and provide a novel PCSD (Pooled Cube and Strip Distillation) continual learning framework to help CMFDFormer handle new tasks. CMFDFormer consists of a MiT (Mix Transformer) backbone network and a PHD (Pluggable Hybrid Decoder) mask prediction network. The MiT backbone network is a Transformer-style network which is adopted on the basis of comprehensive analyses with CNN-style and MLP-style backbones. The PHD network is constructed based on self-correlation computation, hierarchical feature integration, a multi-scale cycle fully-connected block and a mask reconstruction block. The PHD network is applicable to feature extractors of different styles for hierarchical multi-scale information extraction, achieving comparable performance. Last but not least, we propose a PCSD continual learning framework to improve the forgery detectability and avoid catastrophic forgetting when handling new tasks. Our continual learning framework restricts intermediate features from the PHD network, and takes advantage of both cube pooling and strip pooling. Extensive experiments on publicly available datasets demonstrate the good performance of CMFDFormer and the effectiveness of the PCSD continual learning framework.

cs.CV↗

arXiv2026

Progressive-View On-Policy Distillation for Regional-to-Global Transfer in Multimodal LLMs

Regional-to-global distillation uses crop-conditioned guidance to improve full-image understanding. The challenge is to effectively transfer the teacher's crop-based advantage to the student's full-image inference. We propose progressive-view on-policy distillation (PVD), which shifts the student's view distribution from the crop toward the full image through an intermediate aspect-preserving padded crop. The padded crop preserves regional content while matching the full image's visual-token grid. Across stages, the view mixture assigns increasing probability to the full image. A lightweight regional-advantage weighting reallocates token-level supervision using the crop-conditioned teacher-student log-probability gap. Evaluated under each sampled input, it applies mild reweighting when the gap is small and emphasizes higher-gap tokens when the gap widens. A Jensen-Shannon metric decomposition interprets this schedule as a shift from matched-input imitation toward the deployment objective. Across benchmarks spanning perception, visual mathematics and general multimodal question answering, PVD-full reaches an average accuracy of 77.51 over three seeds, improving on the reward-free distillation baseline by 2.01 points and on its reward-matched variant by 1.00 point. In the reward-free setting, PVD-distill still gains 1.16 points.

cs.CV↗

arXiv2026

ALLOT: Budgeted Hybrid-Memory Routing for Knowledge Updates in LLMs

For large language models (LLMs), parametric adaptation is costly when retrieval already suffices. We introduce ALLOT, a hybrid-memory routing framework that separates learned write priority from a hard parametric budget. A memory-aware router combines frozen text representations, retrieval confidence, and relation metadata; a single ranking supports multiple write budgets while preserving all facts in external memory. On CounterFact with Qwen3-4B, ALLOT reaches 0.760 accuracy at a 20% parametric-write budget and recovers 78.4% of the budget-matched oracle gain, with 80% fewer parametric writes than dual-writing every fact. At this budget, jointly adding retrieval and relation features to text improves normalized oracle gain by 6.2 percentage points. Complementary Qwen3-0.6B shared-store results achieve dual-write-level accuracy with 6-14.5% parametric writes, and cross-benchmark transfer retains approximately 88% of in-domain gain. These results support allocating adaptation capacity according to its incremental value rather than treating every factual update as an equally valuable training target.

cs.AI↗

arXiv2026

PRIME: Frequency-Aware Unified Pruning and Merging for MoE Compression

Mixture-of-Experts (MoE) models have emerged as an effective paradigm for scaling large language models, yet their massive parameter footprint poses severe memory challenges for deployment. Existing compression methods typically treat expert pruning and merging as disjoint operations and rely on greedy heuristics, often leading to irreversible information loss or representational pollution. In this paper, we propose PRIME, a unified optimization framework for MoE compression via adaptive PRunIng and MErging. PRIME casts the discrete prune-merge assignment step as a global Integer Linear Programming over a frequency-weighted surrogate objective, jointly selecting pruning and merging decisions while accounting for expert usage. We motivate the frequency-aware objective through an error-bound analysis under fixed representative assignments, and show that the formulation favors high-frequency experts as centroids, promoting routing stability. Extensive experiments on Qwen3-30B-A3B and Qwen3.5-122B-A10B demonstrate that PRIME achieves strong accuracy retention across knowledge, coding, math, and long-context tasks, while serving benchmarks of merged models quantify the memory and throughput benefits of expert compression in vLLM.

cs.LG↗

arXiv2026

WorldMark: A Plug-and-Play World Knowledge Interface for Cross-Host Language Model Watermarking

Watermarking traces the provenance of text produced by large language models by embedding statistically detectable signals during decoding. Existing schemes fall into logits-based, sampling-based, entropy-aware, and adaptive-strength families, yet all of them place watermark signals according to local token statistics. In the open-ended text-generation settings evaluated in this work, local statistics may provide insufficient guidance for placing robust watermark signals. We introduce WorldMark, a plug-and-play interface that uses World Knowledge Memory (WKM) to organize semantic and episodic knowledge in a memory graph, converts the retrieved knowledge into a token-level knowledge saliency score, and adjusts the strength of a host watermark through Asymmetric Knowledge Modulation (AKM). WorldMark requires no backbone retraining and introduces no additional detector-side model or parameter. On the primary C4 evaluation, the complete WorldMark interface improves clean and attacked detection across three adaptive-strength host variants while slightly reducing perplexity. Additional pilot experiments on C4 and OpenGen show that direct memory conditioning transfers across multiple watermark families but can be unstable without saliency-aware modulation. WorldMark requires no additional detector-side model or parameter and introduces negligible overhead under the primary protocol.

cs.CR↗

arXiv2025

MF-Speech: Achieving Fine-Grained and Compositional Control in Speech Generation via Factor Disentanglement

Generating expressive and controllable human speech is one of the core goals of generative artificial intelligence, but its progress has long been constrained by two fundamental challenges: the deep entanglement of speech factors and the coarse granularity of existing control mechanisms. To overcome these challenges, we have proposed a novel framework called MF-Speech, which consists of two core components: MF-SpeechEncoder and MF-SpeechGenerator. MF-SpeechEncoder acts as a factor purifier, adopting a multi-objective optimization strategy to decompose the original speech signal into highly pure and independent representations of content, timbre, and emotion. Subsequently, MF-SpeechGenerator functions as a conductor, achieving precise, composable and fine-grained control over these factors through dynamic fusion and Hierarchical Style Adaptive Normalization (HSAN). Experiments demonstrate that in the highly challenging multi-factor compositional speech generation task, MF-Speech significantly outperforms current state-of-the-art methods, achieving a lower word error rate (WER=4.67%), superior style control (SECS=0.5685, Corr=0.68), and the highest subjective evaluation scores(nMOS=3.96, sMOS_emotion=3.86, sMOS_style=3.78). Furthermore, the learned discrete factors exhibit strong transferability, demonstrating their significant potential as a general-purpose speech representation.

cs.SD↗

arXiv2025

One-Bit Model Aggregation for Differentially Private and Byzantine-Robust Personalized Federated Learning

As the scale of federated learning (FL) systems expands, their inherent performance limitations like communication overhead, Byzantine vulnerability, and privacy leakage have become increasingly critical. This paper considers a personalized FL framework based on model regularization, and proposes a model aggregation algorithm named PRoBit+ to concurrently overcome these limitations. PRoBit+ employs one-bit stochastic quantization and maximum likelihood estimation for parameter aggregation, and dynamically adjusts the step size of parameter updates, improving training stability of deep neural networks under low communication overhead and heterogeneous data distributions. PRoBit+'s statistical analysis is then conducted and its Byzantine robustness is proved. The $(ε,0)$-differential privacy and a convergence upper bound of the PRoBit+ based FL are also theoretically established in heterogeneous contexts. The analysis illustrates the trade-off among transmission accuracy, security guarantees, and convergence rates, and also indicates that the performance degradation caused by transmission errors and privacy protection can be progressively eliminated at a rate of $\mathcal{O}(1/M)$ as the number of uploading clients $M$ increases. Comprehensive numerical experiments are conducted to assess PRoBit+ in comparison to benchmark methods across different Byzantine attacks and varying proportions of malicious clients. The experimental results demonstrate that PRoBit+ exhibits improved Byzantine robustness over existing bit-based transmission schemes, minimal performance degradation related to privacy protection, and nearly identical performance to full-precision FedAvg in a secure environment.

cs.DC↗

arXiv2024

Mechanism of O$_2$ influence on the decomposition process of the eco-friendly gas insulating medium C$_4$F$_7$N/CO$_2$

The C$_4$F$_7$N/CO$_2$/O$_2$ gas mixture is the most promising eco-friendly gas insulation medium available. However, there are few studies on the mechanism of the influence of the buffer gas O2 ratio and its role in the decomposition characteristics of C4F7N/CO2. In this paper, based on the ReaxFF reaction molecular dynamics method and density functional theory, a simulation of the thermal decomposition process of the C$_4$F$_7$N/CO$_2$ mixture under different O2 ratios was carried out at temperatures in the range 2000-3000 K. A constructed model of the C4F7N/CO2/O2 mixture reaction system was used that included the possible reaction paths, product distribution characteristics and their generation rates. The calculation results show that the thermal decomposition of C$_4$F$_7$N/CO$_2$/O$_2$ mainly generates species such as CF$_3$, CF$_2$, CF, F, C$_2$F$_5$, C$_2$F$_4$, C$_2$F$_2$, C$_3$F$_7$, C$_2$F$_2$N, C$_3$F$_4$N, CFN, CN, CO, O, and C. Among them, the two particles CF$_2$ and CN are the most abundant. The first decomposition time of C$_4$F$_7$N is advanced by the addition of O$_2$, while the amount of C$_4$F$_7$N decomposed and the generation of major decomposed particles decreases. The addition of 0%-4% of O$_2$ decreases the reaction rate of the main decomposition reaction in the reaction system. Quantum chemical calculations show that the dissociation process occurring from the combination of C$_4$F$_7$N with O atom is more likely to occur compared to the direct dissociation process of C$_4$F$_7$N molecules. The conclusions of this study provide a theoretical basis for the optimization of the application ratio of C$_4$F$_7$N/CO$_2$/O$_2$ and the diagnosis of its equipment operation and maintenance.

physics.comp-ph↗

arXiv2019

TPM: A GPS-based Trajectory Pattern Mining System

With the development of big data and artificial intelligence, the technology of urban computing becomes more mature and widely used. In urban computing, using GPS-based trajectory data to discover urban dense areas, extract similar urban trajectories, predict urban traffic, and solve traffic congestion problems are all important issues. This paper presents a GPS-based trajectory pattern mining system called TPM. Firstly, the TPM can mine urban dense areas via clustering the spatial-temporal data, and automatically generate trajectories after the timing trajectory identification. Mainly, we propose a method for trajectory similarity matching, and similar trajectories can be extracted via the trajectory similarity matching in this system. The TPM can be applied to the trajectory system equipped with the GPS device, such as the vehicle trajectory, the bicycle trajectory, the electronic bracelet trajectory, etc., to provide services for traffic navigation and journey recommendation. Meantime, the system can provide support in the decision for urban resource allocation, urban functional region identification, traffic congestion and so on.

cs.OH↗

arXiv2019

Joint group and residual sparse coding for image compressive sensing

Nonlocal self-similarity and group sparsity have been widely utilized in image compressive sensing (CS). However, when the sampling rate is low, the internal prior information of degraded images may be not enough for accurate restoration, resulting in loss of image edges and details. In this paper, we propose a joint group and residual sparse coding method for CS image recovery (JGRSC-CS). In the proposed JGRSC-CS, patch group is treated as the basic unit of sparse coding and two dictionaries (namely internal and external dictionaries) are applied to exploit the sparse representation of each group simultaneously. The internal self-adaptive dictionary is used to remove artifacts, and an external Gaussian Mixture Model (GMM) dictionary, learned from clean training images, is used to enhance details and texture. To make the proposed method effective and robust, the split Bregman method is adopted to reconstruct the whole image. Experimental results manifest the proposed JGRSC-CS algorithm outperforms existing state-of-the-art methods in both peak signal to noise ratio (PSNR) and visual quality.

cs.CV↗