Video Representation Learning

Current large-scale video datasets focus on general human activity, but lack depth of coverage on fine-grained activities needed to address physical skill learning. We introduce SportSkills, the first large-scale sports dataset geared towards physical skill learning with in-the-wild video. SportSkills has more than 360k instructional videos containing more than 630k visual demonstrations paired with instructional narrations explaining the know-how behind the actions from 55 varied sports. Through a suite of experiments, we show that SportSkills unlocks the ability to understand fine-grained differences between physical actions. Our representation achieves gains of up to 4x with the same model trained on traditional activity-centric datasets. Crucially, building on SportSkills, we introduce the first large-scale task formulation of mistake-conditioned instructional video retrieval, bridging representation learning and actionable feedback generation (e.g., "here's my execution of a skill; which video clip should I watch to improve it?"). Formal evaluations by professional coaches show our retrieval approach significantly advances the ability of video models to personalize visual instructions for a user query.

Video world models trained with Joint Embedding Predictive Architectures (JEPA) acquire rich spatiotemporal representations by predicting masked regions in latent space rather than reconstructing pixels. This removes the visual verification pathway of generative models, creating a structural interpretability gap: the encoder has learned physical structure inaccessible in any inspectable form. Existing probing methods either operate in continuous space without a structured intermediate layer, or attach generative components whose parameters confound attribution of behavior to the encoder. We propose the AI Mother Tongue (AIM) framework as a passive quantization probe: a lightweight, vocabulary-free probe that converts V-JEPA 2 continuous latent vectors into discrete symbol sequences without task-specific supervision or modifying the encoder. Because the encoder is kept completely frozen, any symbolic structure in the AIM codebook is attributable entirely to V-JEPA 2 pre-trained representations -- not to the probe. We evaluate through category-contrast experiments on Kinetics-mini along three physical dimensions: grasp angle, object geometry, and motion temporal structure. AIM symbol distributions differ significantly across all three experiments (chi^2 p < 10^{-4}; MI 0.036--0.117 bits, NMI 1.2--3.9% of the 3-bit maximum; JSD up to 0.342; codebook active ratio 62.5%). The experiments reveal that V-JEPA 2 latent space is markedly compact: diverse action categories share a common representational core, with semantic differences encoded as graded distributional variations rather than categorical boundaries. These results establish Stage 1 of a four-stage roadmap toward an action-conditioned symbolic world model, demonstrating that structured symbolic manifolds are discoverable properties of frozen JEPA latent spaces.

Current instruction-guided video editing models struggle to simultaneously balance precise semantic modifications with faithful motion preservation. While existing approaches rely on injecting explicit external priors (e.g., VLM features or structural conditions) to mitigate these issues, this reliance severely bottlenecks model robustness and generalization. To overcome this limitation, we present SAMA (factorized Semantic Anchoring and Motion Alignment), a framework that factorizes video editing into semantic anchoring and motion modeling. First, we introduce Semantic Anchoring, which establishes a reliable visual anchor by jointly predicting semantic tokens and video latents at sparse anchor frames, enabling purely instruction-aware structural planning. Second, Motion Alignment pre-trains the same backbone on motion-centric video restoration pretext tasks (cube inpainting, speed perturbation, and tube shuffle), enabling the model to internalize temporal dynamics directly from raw videos. SAMA is optimized with a two-stage pipeline: a factorized pre-training stage that learns inherent semantic-motion representations without paired video-instruction editing data, followed by supervised fine-tuning on paired editing data. Remarkably, the factorized pre-training alone already yields strong zero-shot video editing ability, validating the proposed factorization. SAMA achieves state-of-the-art performance among open-source models and is competitive with leading commercial systems (e.g., Kling-Omni). Code, models, and datasets will be released.

Neural Representations for Videos (NeRV) encode entire video sequences within neural network parameters, offering an alternative paradigm to conventional video codecs. However, the convolutional decoder of NeRV remains computationally expensive and memory intensive, limiting its deployment in resource-constrained environments. This paper proposes LRConv-NeRV, an efficient NeRV variant that replaces selected dense 3x3 convolutional layers with structured low-rank separable convolutions, trained end-to-end within the decoder architecture. By progressively applying low-rank factorization from the largest to earlier decoder stages, LRConv-NeRV enables controllable trade-offs between reconstruction quality and efficiency. Extensive experiments demonstrate that applying LRConv only to the final decoder stage reduces decoder complexity by 68%, from 201.9 to 64.9 GFLOPs, and model size by 9.3%, while incurring negligible quality loss and achieving approximately 9.2% bitrate reduction. Under INT8 post-training quantization, LRConv-NeRV preserves reconstruction quality close to the dense NeRV baseline, whereas more aggressive factorization of early decoder stages leads to disproportionate quality degradation. Compared to existing work under layer-aligned settings, LRConv-NeRV achieves a more favorable efficiency versus quality trade-off, offering substantial GFLOPs and parameter reductions while maintaining higher PSNR/MS-SSIM and improved temporal stability. Temporal flicker analysis using LPIPS further shows that the proposed solution preserves temporal coherence close to the NeRV baseline, results establish LRConv-NeRV as a potential architectural alternative for efficient neural video decoding under low-precision and resource-constrained settings.

Extending language models to video introduces two challenges: representation, where existing methods rely on lossy approximations, and long-context, where caption- or agent-based pipelines collapse video into text and lose visual fidelity. To overcome this, we introduce VideoAtlas, a task-agnostic environment to represent video as a hierarchical grid that is simultaneously lossless, navigable, scalable, caption- and preprocessing-free. An overview of the video is available at a glance, and any region can be recursively zoomed into, with the same visual representation used uniformly for the video, intermediate investigations, and the agent's memory, eliminating lossy text conversion end-to-end. This hierarchical structure ensures access depth grows only logarithmically with video length. For long-context, Recursive Language Models (RLMs) recently offered a powerful solution for long text, but extending them to visual domain requires a structured environment to recurse into, which VideoAtlas provides. VideoAtlas as a Markov Decision Process unlocks Video-RLM: a parallel Master-Worker architecture where a Master coordinates global exploration while Workers concurrently drill into assigned regions to accumulate lossless visual evidence. We demonstrate three key findings: (1)~logarithmic compute growth with video duration, further amplified by a 30-60% multimodal cache hit rate arising from the grid's structural reuse. (2)~environment budgeting, where bounding the maximum exploration depth provides a principled compute-accuracy hyperparameter. (3)~emergent adaptive compute allocation that scales with question granularity. When scaling from 1-hour to 10-hour benchmarks, Video-RLM remains the most duration-robust method with minimal accuracy degradation, demonstrating that structured environment navigation is a viable and scalable paradigm for video understanding.

Tracking Any Point (TAP) has emerged as a fundamental tool for video understanding. Current approaches adapt Vision Foundation Models (VFMs) like DINOv2 via offline finetuning or test-time optimization. However, these VFMs rely on static image pre-training, which is inherently sub-optimal for capturing dense temporal correspondence in videos. To address this, we propose Mask-to-Point (M2P) learning, which leverages rich video object segmentation (VOS) mask annotations to improve VFMs for dense point tracking. Our M2P introduces three new mask-based constraints for weakly-supervised representation learning. First, we propose a local structure consistency loss, which leverages Procrustes analysis to model the cohesive motion of points lying within a local structure, achieving more reliable point-to-point matching learning. Second, we propose a mask label consistency (MLC) loss, which enforces that sampled foreground points strictly match foreground regions across frames. The proposed MLC loss can be regarded as a regularization, which stabilizes training and prevents convergence to trivial solutions. Finally, mask boundary constrain is applied to explicitly supervise boundary points. We show that our weaklysupervised M2P models significantly outperform baseline VFMs with efficient training by using only 3.6K VOS training videos. Notably, M2P achieves 12.8% and 14.6% performance gains over DINOv2-B/14 and DINOv3-B/16 on the TAP-Vid-DAVIS benchmark, respectively. Moreover, the proposed M2P models are used as pre-trained backbones for both test-time optimized and offline fine-tuned TAP tasks, demonstrating its potential to serve as general pre-trained models for point tracking. Code will be made publicly available upon acceptance.

We present V-JEPA 2.1, a family of self-supervised models that learn dense, high-quality visual representations for both images and videos while retaining strong global scene understanding. The approach combines four key components. First, a dense predictive loss uses a masking-based objective in which both visible and masked tokens contribute to the training signal, encouraging explicit spatial and temporal grounding. Second, deep self-supervision applies the self-supervised objective hierarchically across multiple intermediate encoder layers to improve representation quality. Third, multi-modal tokenizers enable unified training across images and videos. Finally, the model benefits from effective scaling in both model capacity and training data. Together, these design choices produce representations that are spatially structured, semantically coherent, and temporally consistent. Empirically, V-JEPA 2.1 achieves state-of-the-art performance on several challenging benchmarks, including 7.71 mAP on Ego4D for short-term object-interaction anticipation and 40.8 Recall@5 on EPIC-KITCHENS for high-level action anticipation, as well as a 20-point improvement in real-robot grasping success rate over V-JEPA-2 AC. The model also demonstrates strong performance in robotic navigation (5.687 ATE on TartanDrive), depth estimation (0.307 RMSE on NYUv2 with a linear probe), and global recognition (77.7 on Something-Something-V2). These results show that V-JEPA 2.1 significantly advances the state of the art in dense visual understanding and world modeling.

Video-based lie detection aims to identify deceptive behaviors from visual cues. Despite recent progress, its core challenge lies in learning sparse yet discriminative representations. Deceptive signals are typically subtle and short-lived, easily overwhelmed by redundant information, while individual and contextual variations introduce strong identity-related noise. To address this issue, we propose GenLie, a Global-Enhanced Lie Detection Network that performs local feature modeling under global supervision. Specifically, sparse and subtle deceptive cues are captured at the local level, while global supervision and optimization ensure robust and discriminative representations by suppressing identity-related noise. Experiments on three public datasets, covering both high- and low-stakes scenarios, show that GenLie consistently outperforms state-of-the-art methods. Source code is available at https://github.com/AliasDictusZ1/GenLie.

Humans develop visual intelligence through perceiving and interacting with their environment - a self-supervised learning process grounded in egocentric experience. Inspired by this, we ask how can artificial systems learn stable object representations from continuous, uncurated first-person videos without relying on manual annotations. This setting poses challenges of separating, recognizing, and persistently tracking objects amid clutter, occlusion, and ego-motion. We propose EgoViT, a unified vision Transformer framework designed to learn stable object representations from unlabeled egocentric video. EgoViT bootstraps this learning process by jointly discovering and stabilizing "proto-objects" through three synergistic mechanisms: (1) Proto-object Learning, which uses intra-frame distillation to form discriminative representations; (2) Depth Regularization, which grounds these representations in geometric structure; and (3) Teacher-Filtered Temporal Consistency, which enforces identity over time. This creates a virtuous cycle where initial object hypotheses are progressively refined into stable, persistent representations. The framework is trained end-to-end on unlabeled first-person videos and exhibits robustness to geometric priors of varied origin and quality. On standard benchmarks, EgoViT achieves +8.0% CorLoc improvement in unsupervised object discovery and +4.8% mIoU improvement in semantic segmentation, demonstrating its potential to lay a foundation for robust visual abstraction in embodied intelligence.

Modern visual agents require representations that are general, causal, and physically structured to operate in real-time streaming environments. However, current vision foundation models remain fragmented, specializing narrowly in image semantic perception, offline temporal modeling, or spatial geometry. This paper introduces OmniStream, a unified streaming visual backbone that effectively perceives, reconstructs, and acts from diverse visual inputs. By incorporating causal spatiotemporal attention and 3D rotary positional embeddings (3D-RoPE), our model supports efficient, frame-by-frame online processing of video streams via a persistent KV-cache. We pre-train OmniStream using a synergistic multi-task framework coupling static and temporal representation learning, streaming geometric reconstruction, and vision-language alignment on 29 datasets. Extensive evaluations show that, even with a strictly frozen backbone, OmniStream achieves consistently competitive performance with specialized experts across image and video probing, streaming geometric reconstruction, complex video and spatial reasoning, as well as robotic manipulation (unseen at training). Rather than pursuing benchmark-specific dominance, our work demonstrates the viability of training a single, versatile vision backbone that generalizes across semantic, spatial, and temporal reasoning, i.e., a more meaningful step toward general-purpose visual understanding for interactive and embodied agents.

The paper explores how video models trained for classification tasks represent nuanced, hidden semantic information that may not affect the final outcome, a key challenge for Trustworthy AI models. Through Explainable and Interpretable AI methods, specifically mechanistic interpretability techniques, the internal circuit responsible for representing the action's outcome is reverse-engineered in a pre-trained video vision transformer, revealing that the "Success vs Failure" signal is computed through a distinct amplification cascade. While there are low-level differences observed from layer 0, the abstract and semantic representation of the outcome is progressively amplified from layers 5 through 11. Causal analysis, primarily using activation patching supported by ablation results, reveals a clear division of labor: Attention Heads act as "evidence gatherers", providing necessary low-level information for partial signal recovery, while MLP Blocks function as robust "concept composers", each of which is the primary driver to generate the "success" signal. This distributed and redundant circuit in the model's internals explains its resilience to simple ablations, demonstrating a core computational pattern for processing human-action outcomes. Crucially, the existence of this sophisticated circuit for representing complex outcomes, even within a model trained only for simple classification, highlights the potential for models to develop forms of 'hidden knowledge' beyond their explicit task, underscoring the need for mechanistic oversight for building genuinely Explainable and Trustworthy AI systems intended for deployment.

Implicit Neural Representations (INRs) have emerged as a promising paradigm for video representation and compression. However, existing multi-scale INR generators often suffer from significant parameter redundancy by stacking independent processing blocks for each scale. Inspired by the principle of scale self-similarity in the generation process, we propose SRNeRV, a novel scale-wise recursive framework that replaces this stacked design with a parameter-efficient shared architecture. The core of our approach is a hybrid sharing scheme derived from decoupling the processing block into a scale-specific spatial mixing module and a scale-invariant channel mixing module. We recursively apply the same shared channel mixing module, which contains the majority of the parameters, across all scales, significantly reducing the model size while preserving the crucial capacity to learn scale-specific spatial patterns. Extensive experiments demonstrate that SRNeRV achieves a significant rate-distortion performance boost, especially in INR-friendly scenarios, validating that our sharing scheme successfully amplifies the core strengths of the INR paradigm.

For video-text retrieval, the use of CLIP has been a de facto choice. Since CLIP provides only image and text encoders, this consensus has led to a biased paradigm that entirely ignores the sound track of videos. While several attempts have been made to reintroduce audio -- typically by incorporating an audio encoder and fusing its output with visual features -- these methods face two challenges: ineffective representation of speech content and suboptimal vision-audio fusion. To address these issues jointly, we propose SAVE, a Speech Aware Video rEpresentation learning method. SAVE improves upon AVIGATE, a SOTA audiovisual method, with a dedicated speech branch for more effective speech embedding. Furthermore, we introduce soft-ALBEF for early vision-audio alignment that facilitates fusion. Extensive experiments on five benchmarks show that SAVE compares favorably against the SOTA, outperforming AVIGATE by +4.1% on MSRVTT-9k, +1.9% on MSRVTT-7k, +2.5% on VATEX, +9.8% on Charades, and +2.1% on LSMDC, in light of the SumR metric.

Self-supervised learning (SSL) has made rapid progress, yet learned features often over-rely on contextual shortcuts-background textures and co-occurrence statistics. While video provides rich temporal variation, dense in-the-wild streams with strong ego-motion create a co-occurrence trap: foreground objects and background context move coherently, encouraging representations to collapse into scene encoders. To address this, we propose VINO (Video-driven Invariance for Non-Contextual Objects), a teacher-student framework that learns robust image encoders from dense video by imposing a structural information bottleneck. Using a class-agnostic structural prior solely to generate views-not as semantic pseudo-labels-VINO forms an asymmetric distillation problem. The teacher predicts from a foreground-union view with the background suppressed, while the student observes object-conditioned scene views that retain surrounding context but remove competing instances. Matching these targets via masked distillation makes background cues unreliable, pushing the representation toward object-centric invariances. We further enforce temporal object permanence via teacher-anchored cross-time distillation over track-matched objects, and stabilize part-to-whole consistency with mask-guided local views. Through attention visualization and unsupervised object discovery on PASCAL VOC, we demonstrate that VINO effectively disentangles foreground from background. Pretrained on the dense Walking Tours Venice video, VINO achieves 34.8 CorLoc, yielding highly focused, shape-biased representations that substantially outperform prior dense-video and motion-guided SSL baselines.

Tokenization in video models, typically through patchification, generates an excessive and redundant number of tokens. This severely limits video efficiency and scalability. While recent trajectory-based tokenizers offer a promising solution by decoupling video duration from token count, they rely on complex external segmentation and tracking pipelines that are slow and task-agnostic. We propose TrajTok, an end-to-end video tokenizer module that is fully integrated and co-trained with video models for a downstream objective, dynamically adapting its token granularity to semantic complexity, independent of video duration. TrajTok contains a unified segmenter that performs implicit clustering over pixels in both space and time to directly produce object trajectories in a single forward pass. By prioritizing downstream adaptability over pixel-perfect segmentation fidelity, TrajTok is lightweight and efficient, yet empirically improves video understanding performance. With TrajTok, we implement a video CLIP model trained from scratch (TrajViT2). It achieves the best accuracy at scale across both classification and retrieval benchmarks, while maintaining efficiency comparable to the best token-merging methods. TrajTok also proves to be a versatile component beyond its role as a tokenizer. We show that it can be seamlessly integrated as either a probing head for pretrained visual features (TrajAdapter) or an alignment connector in vision-language models (TrajVLM) with especially strong performance in long-video reasoning.

Ego-centric driving videos available online provide an abundant source of visual data for autonomous driving, yet their lack of annotations makes it difficult to learn representations that capture both semantic structure and 3D geometry. Recent advances in large feedforward spatial models demonstrate that point maps and ego-motion can be inferred in a single forward pass, suggesting a promising direction for scalable driving perception. We therefore propose a label-free, teacher-guided framework for learning autonomous driving representations directly from unposed videos. Unlike prior self-supervised approaches that focus primarily on frame-to-frame consistency, we posit that safe and reactive driving depends critically on temporal context. To this end, we leverage a feedforward architecture equipped with a lightweight autoregressive module, trained using multi-modal supervisory signals that guide the model to jointly predict current and future point maps, camera poses, semantic segmentation, and motion masks. Multi-modal teachers provide sequence-level pseudo-supervision, enabling LFG to learn a unified pseudo-4D representation from raw YouTube videos without poses, labels, or LiDAR. The resulting encoder not only transfers effectively to downstream autonomous driving planning on the NAVSIM benchmark, surpassing multi-camera and LiDAR baselines with only a single monocular camera, but also yields strong performance when evaluated on a range of semantic, geometric, and qualitative motion prediction tasks. These geometry and motion-aware features position LFG as a compelling video-centric foundation model for autonomous driving.

We present Xray-Visual, a unified vision model architecture for large-scale image and video understanding trained on industry-scale social media data. Our model leverages over 15 billion curated image-text pairs and 10 billion video-hashtag pairs from Facebook and Instagram, employing robust data curation pipelines that incorporate balancing and noise suppression strategies to maximize semantic diversity while minimizing label noise. We introduce a three-stage training pipeline that combines self-supervised MAE, semi-supervised hashtag classification, and CLIP-style contrastive learning to jointly optimize image and video modalities. Our architecture builds on a Vision Transformer backbone enhanced with efficient token reorganization (EViT) for improved computational efficiency. Extensive experiments demonstrate that Xray-Visual achieves state-of-the-art performance across diverse benchmarks, including ImageNet for image classification, Kinetics and HMDB51 for video understanding, and MSCOCO for cross-modal retrieval. The model exhibits strong robustness to domain shift and adversarial perturbations. We further demonstrate that integrating large language models as text encoders (LLM2CLIP) significantly enhances retrieval performance and generalization capabilities, particularly in real-world environments. Xray-Visual establishes new benchmarks for scalable, multimodal vision models, while maintaining superior accuracy and computational efficiency.

While multimodal large language models (MLLMs) have shown remarkable success across a wide range of tasks, long-form video understanding remains a significant challenge. In this study, we focus on video understanding by MLLMs. This task is challenging because processing a full stream of RGB frames is computationally intractable and highly redundant, as self-attention have quadratic complexity with sequence length. In this paper, we propose ReMoRa, a video MLLM that processes videos by operating directly on their compressed representations. A sparse set of RGB keyframes is retained for appearance, while temporal dynamics are encoded as a motion representation, removing the need for sequential RGB frames. These motion representations act as a compact proxy for optical flow, capturing temporal dynamics without full frame decoding. To refine the noise and low fidelity of block-based motions, we introduce a module to denoise and generate a fine-grained motion representation. Furthermore, our model compresses these features in a way that scales linearly with sequence length. We demonstrate the effectiveness of ReMoRa through extensive experiments across a comprehensive suite of long-video understanding benchmarks. ReMoRa outperformed baseline methods on multiple challenging benchmarks, including LongVideoBench, NExT-QA, and MLVU.

Temporal information has long been considered to be essential for perception. While there is extensive research on the role of image information for perceptual tasks, the role of the temporal dimension remains less well understood: What can we learn about the world from long-term motion information? What properties does long-term motion information have for visual learning? We leverage recent success in point-track estimation, which offers an excellent opportunity to learn temporal representations and experiment on a variety of perceptual tasks. We draw 3 clear lessons: 1) Long-term motion representations contain information to understand actions, but also objects, materials, and spatial information, often even better than images. 2) Long-term motion representations generalize far better than image representations in low-data settings and in zero-shot tasks. 3) The very low dimensionality of motion information makes motion representations a better trade-off between GFLOPs and accuracy than standard video representations, and used together they achieve higher performance than video representations alone. We hope these insights will pave the way for the design of future models that leverage the power of long-term motion information for perception.

While foundation models have advanced surgical video analysis, current approaches rely predominantly on pixel-level reconstruction objectives that waste model capacity on low-level visual details, such as smoke, specular reflections, and fluid motion, rather than semantic structures essential for surgical understanding. We present SurgMotion, a video-native foundation model that shifts the learning paradigm from pixel-level reconstruction to latent motion prediction. Built on the Video Joint Embedding Predictive Architecture (V-JEPA), SurgMotion introduces three key technical innovations tailored to surgical videos: (1) motion-guided latent masked prediction to prioritize semantically meaningful regions, (2) spatiotemporal affinity self-distillation to enforce relational consistency, and (3) spatiotemporal feature diversity regularization (SFDR) to prevent representation collapse in texture-sparse surgical scenes. To enable large-scale pretraining, we curate SurgMotion-15M, the largest surgical video dataset to date, comprising 3,658 hours of video from 50 sources across 13 anatomical regions. Extensive experiments across 17 benchmarks demonstrate that SurgMotion significantly outperforms state-of-the-art methods on surgical workflow recognition, achieving 14.6 percent improvement in F1 score on EgoSurgery and 10.3 percent on PitVis; on action triplet recognition with 39.54 percent mAP-IVT on CholecT50; as well as on skill assessment, polyp segmentation, and depth estimation. These results establish SurgMotion as a new standard for universal, motion-oriented surgical video understanding.

End-to-end autonomous driving increasingly leverages self-supervised video pretraining to learn transferable planning representations. However, pretraining video world models for scene understanding has so far brought only limited improvements. This limitation is compounded by the inherent ambiguity of driving: each scene typically provides only a single human trajectory, making it difficult to learn multimodal behaviors. In this work, we propose Drive-JEPA, a framework that integrates Video Joint-Embedding Predictive Architecture (V-JEPA) with multimodal trajectory distillation for end-to-end driving. First, we adapt V-JEPA for end-to-end driving, pretraining a ViT encoder on large-scale driving videos to produce predictive representations aligned with trajectory planning. Second, we introduce a proposal-centric planner that distills diverse simulator-generated trajectories alongside human trajectories, with a momentum-aware selection mechanism to promote stable and safe behavior. When evaluated on NAVSIM, the V-JEPA representation combined with a simple transformer-based decoder outperforms prior methods by 3 PDMS in the perception-free setting. The complete Drive-JEPA framework achieves 93.3 PDMS on v1 and 87.8 EPDMS on v2, setting a new state-of-the-art.

Short-form video moderation is increasingly pushed toward edge and privacy-sensitive settings, where users may intend videos for a limited audience, such as friends or private groups, but sending raw clips to a central server can broaden exposure, consume bandwidth, and add moderation latency. Federated learning can keep videos on device, but unprotected model updates may still leak information, and full-video backbones are expensive to communicate. We present FedVideoMAE, a privacy-preserving federated framework for violence detection that adapts a frozen VideoMAE backbone with lightweight LoRA and prompt parameters. Each training round combines self-supervised masked video reconstruction with client-side differential privacy and pairwise masked aggregation (SA) of adapter updates. Violence labels are held out from federation and used only for downstream evaluation, separating private representation learning from supervised assessment. On RWF-2000, exchanging 5,518,848 trainable parameters instead of the 156,371,328-parameter instantiated pretraining state gives a 28.3x model-state payload ratio. FedVideoMAE reaches 77.25% test accuracy without DP or SA, while accuracy under DP+SA remains in the 65.25-66.00% range. Transfer experiments on RLVS and binary UCF-Crime show similar behavior. These results characterize the privacy-utility trade-off for edge video moderation. Code is available at: https://github.com/zyt-599/FedVideoMAE

Masked modeling has emerged as a robust self-supervised learning framework. However, most methods rely on random masking, which disregards the structural properties of different data modalities. To align with the spatiotemporal and spectral characteristics of video and audio data, we introduce a structured noise-based masking approach. By filtering white noise into different color noise distributions, we generate structured masks that capture modality-specific patterns without requiring handcrafted heuristics or access to the data. Our approach enhances masked video and audio modeling frameworks without any additional computational cost. Experiments show that structured noise masking consistently outperforms random masking, underscoring the value of modality-aware masking strategies for representation learning.