Changick Kim

Sangmin Woo, Sumin Lee, Yeonju Park et al.

Standard multi-modal models assume the use of the same modalities in training and inference stages. However, in practice, the environment in which multi-modal models operate may not satisfy such assumption. As such, their performances degrade drastically if any modality is missing in the inference stage. We ask: how can we train a model that is robust to missing modalities? This paper seeks a set of good practices for multi-modal action recognition, with a particular interest in circumstances where some modalities are not available at an inference time. First, we study how to effectively regularize the model during training (e.g., data augmentation). Second, we investigate on fusion methods for robustness to missing modalities: we find that transformer-based fusion shows better robustness for missing modality than summation or concatenation. Third, we propose a simple modular network, ActionMAE, which learns missing modality predictive coding by randomly dropping modality features and tries to reconstruct them with the remaining modality features. Coupling these good practices, we build a model that is not only effective in multi-modal action recognition but also robust to modality missing. Our model achieves the state-of-the-arts on multiple benchmarks and maintains competitive performances even in missing modality scenarios. Codes are available at https://github.com/sangminwoo/ActionMAE.

Jinyoung Park, Minseok Son, Seungju Cho et al.

This paper presents a solution to the Weather4cast 2022 Challenge Stage 2. The goal of the challenge is to forecast future high-resolution rainfall events obtained from ground radar using low-resolution multiband satellite images. We suggest a solution that performs data preprocessing appropriate to the challenge and then predicts rainfall movies using a novel RainUNet. RainUNet is a hierarchical U-shaped network with temporal-wise separable block (TS block) using a decoupled large kernel 3D convolution to improve the prediction performance. Various evaluation metrics show that our solution is effective compared to the baseline method. The source codes are available at https://github.com/jinyxp/Weather4cast-2022

Junyoung Byun, Seungju Cho, Myung-Joon Kwon et al.

The transferability of adversarial examples allows the deception on black-box models, and transfer-based targeted attacks have attracted a lot of interest due to their practical applicability. To maximize the transfer success rate, adversarial examples should avoid overfitting to the source model, and image augmentation is one of the primary approaches for this. However, prior works utilize simple image transformations such as resizing, which limits input diversity. To tackle this limitation, we propose the object-based diverse input (ODI) method that draws an adversarial image on a 3D object and induces the rendered image to be classified as the target class. Our motivation comes from the humans' superior perception of an image printed on a 3D object. If the image is clear enough, humans can recognize the image content in a variety of viewing conditions. Likewise, if an adversarial example looks like the target class to the model, the model should also classify the rendered image of the 3D object as the target class. The ODI method effectively diversifies the input by leveraging an ensemble of multiple source objects and randomizing viewing conditions. In our experimental results on the ImageNet-Compatible dataset, this method boosts the average targeted attack success rate from 28.3% to 47.0% compared to the state-of-the-art methods. We also demonstrate the applicability of the ODI method to adversarial examples on the face verification task and its superior performance improvement. Our code is available at https://github.com/dreamflake/ODI.

Inyong Koo, Inyoung Lee, Se-Ho Kim et al.

One of the main challenges in LiDAR-based 3D object detection is that the sensors often fail to capture the complete spatial information about the objects due to long distance and occlusion. Two-stage detectors with point cloud completion approaches tackle this problem by adding more points to the regions of interest (RoIs) with a pre-trained network. However, these methods generate dense point clouds of objects for all region proposals, assuming that objects always exist in the RoIs. This leads to the indiscriminate point generation for incorrect proposals as well. Motivated by this, we propose Point Generation R-CNN (PG-RCNN), a novel end-to-end detector that generates semantic surface points of foreground objects for accurate detection. Our method uses a jointly trained RoI point generation module to process the contextual information of RoIs and estimate the complete shape and displacement of foreground objects. For every generated point, PG-RCNN assigns a semantic feature that indicates the estimated foreground probability. Extensive experiments show that the point clouds generated by our method provide geometrically and semantically rich information for refining false positive and misaligned proposals. PG-RCNN achieves competitive performance on the KITTI benchmark, with significantly fewer parameters than state-of-the-art models. The code is available at https://github.com/quotation2520/PG-RCNN.

Hansol Kim, Youngjun Kwak, Minyoung Jung et al.

Federated learning (FL) is a promising approach for enhancing data privacy preservation, particularly for authentication systems. However, limited round communications, scarce representation, and scalability pose significant challenges to its deployment, hindering its full potential. In this paper, we propose 'ProtoFL', Prototypical Representation Distillation based unsupervised Federated Learning to enhance the representation power of a global model and reduce round communication costs. Additionally, we introduce a local one-class classifier based on normalizing flows to improve performance with limited data. Our study represents the first investigation of using FL to improve one-class classification performance. We conduct extensive experiments on five widely used benchmarks, namely MNIST, CIFAR-10, CIFAR-100, ImageNet-30, and Keystroke-Dynamics, to demonstrate the superior performance of our proposed framework over previous methods in the literature.

Sumin Lee, Sangmin Woo, Yeonju Park et al.

In multi-modal action recognition, it is important to consider not only the complementary nature of different modalities but also global action content. In this paper, we propose a novel network, named Modality Mixer (M-Mixer) network, to leverage complementary information across modalities and temporal context of an action for multi-modal action recognition. We also introduce a simple yet effective recurrent unit, called Multi-modal Contextualization Unit (MCU), which is a core component of M-Mixer. Our MCU temporally encodes a sequence of one modality (e.g., RGB) with action content features of other modalities (e.g., depth, IR). This process encourages M-Mixer to exploit global action content and also to supplement complementary information of other modalities. As a result, our proposed method outperforms state-of-the-art methods on NTU RGB+D 60, NTU RGB+D 120, and NW-UCLA datasets. Moreover, we demonstrate the effectiveness of M-Mixer by conducting comprehensive ablation studies.

Byeongjun Park, Hyojun Go, Changick Kim

Creating novel views from a single image has achieved tremendous strides with advanced autoregressive models, as unseen regions have to be inferred from the visible scene contents. Although recent methods generate high-quality novel views, synthesizing with only one explicit or implicit 3D geometry has a trade-off between two objectives that we call the "seesaw" problem: 1) preserving reprojected contents and 2) completing realistic out-of-view regions. Also, autoregressive models require a considerable computational cost. In this paper, we propose a single-image view synthesis framework for mitigating the seesaw problem while utilizing an efficient non-autoregressive model. Motivated by the characteristics that explicit methods well preserve reprojected pixels and implicit methods complete realistic out-of-view regions, we introduce a loss function to complement two renderers. Our loss function promotes that explicit features improve the reprojected area of implicit features and implicit features improve the out-of-view area of explicit features. With the proposed architecture and loss function, we can alleviate the seesaw problem, outperforming autoregressive-based state-of-the-art methods and generating an image $\approx$100 times faster. We validate the efficiency and effectiveness of our method with experiments on RealEstate10K and ACID datasets.

Youngjun Kwak, Minyoung Jung, Hunjae Yoo et al.

Previous anti-spoofing methods have used either pseudo maps or user-defined labels, and the performance of each approach depends on the accuracy of the third party networks generating pseudo maps and the way in which the users define the labels. In this paper, we propose a liveness score-based regression network for overcoming the dependency on third party networks and users. First, we introduce a new labeling technique, called pseudo-discretized label encoding for generating discretized labels indicating the amount of information related to real images. Secondly, we suggest the expected liveness score based on a regression network for training the difference between the proposed supervision and the expected liveness score. Finally, extensive experiments were conducted on four face anti-spoofing benchmarks to verify our proposed method on both intra-and cross-dataset tests. The experimental results show our approach outperforms previous methods.

Jeongsoo Kim, Sangmin Woo, Byeongjun Park et al.

Camera traps, unmanned observation devices, and deep learning-based image recognition systems have greatly reduced human effort in collecting and analyzing wildlife images. However, data collected via above apparatus exhibits 1) long-tailed and 2) open-ended distribution problems. To tackle the open-set long-tailed recognition problem, we propose the Temporal Flow Mask Attention Network that comprises three key building blocks: 1) an optical flow module, 2) an attention residual module, and 3) a meta-embedding classifier. We extract temporal features of sequential frames using the optical flow module and learn informative representation using attention residual blocks. Moreover, we show that applying the meta-embedding technique boosts the performance of the method in open-set long-tailed recognition. We apply this method on a Korean Demilitarized Zone (DMZ) dataset. We conduct extensive experiments, and quantitative and qualitative analyses to prove that our method effectively tackles the open-set long-tailed recognition problem while being robust to unknown classes.

Minki Jeong, Changick Kim

Real-world data often have a long-tailed distribution, where the number of samples per class is not equal over training classes. The imbalanced data form a biased feature space, which deteriorates the performance of the recognition model. In this paper, we propose a novel long-tailed recognition method to balance the latent feature space. First, we introduce a MixUp-based data augmentation technique to reduce the bias of the long-tailed data. Furthermore, we propose a new supervised contrastive learning method, named Supervised contrastive learning on Mixed Classes (SMC), for blended images. SMC creates a set of positives based on the class labels of the original images. The combination ratio of positives weights the positives in the training loss. SMC with the class-mixture-based loss explores more diverse data space, enhancing the generalization capability of the model. Extensive experiments on various benchmarks show the effectiveness of our one-stage training method.

Byeongjun Park, Sangmin Woo, Hyojun Go et al.

Diffusion models generate highly realistic images by learning a multi-step denoising process, naturally embodying the principles of multi-task learning (MTL). Despite the inherent connection between diffusion models and MTL, there remains an unexplored area in designing neural architectures that explicitly incorporate MTL into the framework of diffusion models. In this paper, we present Denoising Task Routing (DTR), a simple add-on strategy for existing diffusion model architectures to establish distinct information pathways for individual tasks within a single architecture by selectively activating subsets of channels in the model. What makes DTR particularly compelling is its seamless integration of prior knowledge of denoising tasks into the framework: (1) Task Affinity: DTR activates similar channels for tasks at adjacent timesteps and shifts activated channels as sliding windows through timesteps, capitalizing on the inherent strong affinity between tasks at adjacent timesteps. (2) Task Weights: During the early stages (higher timesteps) of the denoising process, DTR assigns a greater number of task-specific channels, leveraging the insight that diffusion models prioritize reconstructing global structure and perceptually rich contents in earlier stages, and focus on simple noise removal in later stages. Our experiments reveal that DTR not only consistently boosts diffusion models' performance across different evaluation protocols without adding extra parameters but also accelerates training convergence. Finally, we show the complementarity between our architectural approach and existing MTL optimization techniques, providing a more complete view of MTL in the context of diffusion training. Significantly, by leveraging this complementarity, we attain matched performance of DiT-XL using the smaller DiT-L with a reduction in training iterations from 7M to 2M.

Sangmin Woo, So-Yeong Jeon, Jinyoung Park et al.

We introduce Sketch-based Video Object Localization (SVOL), a new task aimed at localizing spatio-temporal object boxes in video queried by the input sketch. We first outline the challenges in the SVOL task and build the Sketch-Video Attention Network (SVANet) with the following design principles: (i) to consider temporal information of video and bridge the domain gap between sketch and video; (ii) to accurately identify and localize multiple objects simultaneously; (iii) to handle various styles of sketches; (iv) to be classification-free. In particular, SVANet is equipped with a Cross-modal Transformer that models the interaction between learnable object tokens, query sketch, and video through attention operations, and learns upon a per-frame set matching strategy that enables frame-wise prediction while utilizing global video context. We evaluate SVANet on a newly curated SVOL dataset. By design, SVANet successfully learns the mapping between the query sketches and video objects, achieving state-of-the-art results on the SVOL benchmark. We further confirm the effectiveness of SVANet via extensive ablation studies and visualizations. Lastly, we demonstrate its transfer capability on unseen datasets and novel categories, suggesting its high scalability in real-world applications.

Seokil Ham, Jaehyuk Jang, Wonjun Lee et al.

Fine-tuning-as-a-Service (FaaS) enables personalization of large language models (LLMs), but it can weaken safety-alignment under harmful fine-tuning attacks. Recent work has shown that activating harmful-behavior modules during fine-tuning can prevent models from learning undesired behaviors, but its mechanism remains unclear. In this paper, we revisit temporary jailbreaking as a defense against harmful fine-tuning and provide a gradient-level analysis showing that it saturates safety-degrading gradients while preserving benign task-relevant gradients. Based on this insight, we propose a Buffer-and-Reinforce fine-tuning framework that buffers harmful updates during user fine-tuning and reinforces safety after adaptation. Specifically, BufferLoRA induces temporary jailbreaking as a removable adapter to reduce harmful updates during user fine-tuning. After adaptation, ReinforceLoRA, trained to recover refusal behavior under the temporarily jailbroken state, is integrated with UserLoRA via QR decomposition-based merging to reinforce safety while preserving user-task performance. Extensive experiments show that our framework achieves superior safety and utility with no additional safety data during user fine-tuning and minimal computational cost.

Jinyoung Park, Hee-Seon Kim, Kangwook Ko et al.

We introduce VideoMamba, a novel adaptation of the pure Mamba architecture, specifically designed for video recognition. Unlike transformers that rely on self-attention mechanisms leading to high computational costs by quadratic complexity, VideoMamba leverages Mamba's linear complexity and selective SSM mechanism for more efficient processing. The proposed Spatio-Temporal Forward and Backward SSM allows the model to effectively capture the complex relationship between non-sequential spatial and sequential temporal information in video. Consequently, VideoMamba is not only resource-efficient but also effective in capturing long-range dependency in videos, demonstrated by competitive performance and outstanding efficiency on a variety of video understanding benchmarks. Our work highlights the potential of VideoMamba as a powerful tool for video understanding, offering a simple yet effective baseline for future research in video analysis.

Hee-Seon Kim, Minji Son, Minbeom Kim et al.

As video analysis using deep learning models becomes more widespread, the vulnerability of such models to adversarial attacks is becoming a pressing concern. In particular, Universal Adversarial Perturbation (UAP) poses a significant threat, as a single perturbation can mislead deep learning models on entire datasets. We propose a novel video UAP using image data and image model. This enables us to take advantage of the rich image data and image model-based studies available for video applications. However, there is a challenge that image models are limited in their ability to analyze the temporal aspects of videos, which is crucial for a successful video attack. To address this challenge, we introduce the Breaking Temporal Consistency (BTC) method, which is the first attempt to incorporate temporal information into video attacks using image models. We aim to generate adversarial videos that have opposite patterns to the original. Specifically, BTC-UAP minimizes the feature similarity between neighboring frames in videos. Our approach is simple but effective at attacking unseen video models. Additionally, it is applicable to videos of varying lengths and invariant to temporal shifts. Our approach surpasses existing methods in terms of effectiveness on various datasets, including ImageNet, UCF-101, and Kinetics-400.

Byeongjun Park, Changick Kim

Dynamic radiance fields have emerged as a promising approach for generating novel views from a monocular video. However, previous methods enforce the geometric consistency to dynamic radiance fields only between adjacent input frames, making it difficult to represent the global scene geometry and degenerates at the viewpoint that is spatio-temporally distant from the input camera trajectory. To solve this problem, we introduce point-based dynamic radiance fields (\textbf{Point-DynRF}), a novel framework where the global geometric information and the volume rendering process are trained by neural point clouds and dynamic radiance fields, respectively. Specifically, we reconstruct neural point clouds directly from geometric proxies and optimize both radiance fields and the geometric proxies using our proposed losses, allowing them to complement each other. We validate the effectiveness of our method with experiments on the NVIDIA Dynamic Scenes Dataset and several causally captured monocular video clips.

Inyong Koo, yeeun Seong, Minseok Son et al.

Audio-visual emotion recognition (AVER) methods typically fuse utterance-level features, and even frame-level attention models seldom address the frame-rate mismatch across modalities. In this paper, we propose a Transformer-based framework focusing on the temporal alignment of multimodal features. Our design employs a multimodal self-attention encoder that simultaneously captures intra- and inter-modal dependencies within a shared feature space. To address heterogeneous sampling rates, we incorporate Temporally-aligned Rotary Position Embeddings (TaRoPE), which implicitly synchronize audio and video tokens. Furthermore, we introduce a Cross-Temporal Matching (CTM) loss that enforces consistency among temporally proximate pairs, guiding the encoder toward better alignment. Experiments on CREMA-D and RAVDESS datasets demonstrate consistent improvements over recent baselines, suggesting that explicitly addressing frame-rate mismatch helps preserve temporal cues and enhances cross-modal fusion.

Muhammad Adi Nugroho, Sangmin Woo, Sumin Lee et al.

Deep learning has made significant strides in video understanding tasks, but the computation required to classify lengthy and massive videos using clip-level video classifiers remains impractical and prohibitively expensive. To address this issue, we propose Audio-Visual Glance Network (AVGN), which leverages the commonly available audio and visual modalities to efficiently process the spatio-temporally important parts of a video. AVGN firstly divides the video into snippets of image-audio clip pair and employs lightweight unimodal encoders to extract global visual features and audio features. To identify the important temporal segments, we use an Audio-Visual Temporal Saliency Transformer (AV-TeST) that estimates the saliency scores of each frame. To further increase efficiency in the spatial dimension, AVGN processes only the important patches instead of the whole images. We use an Audio-Enhanced Spatial Patch Attention (AESPA) module to produce a set of enhanced coarse visual features, which are fed to a policy network that produces the coordinates of the important patches. This approach enables us to focus only on the most important spatio-temporally parts of the video, leading to more efficient video recognition. Moreover, we incorporate various training techniques and multi-modal feature fusion to enhance the robustness and effectiveness of our AVGN. By combining these strategies, our AVGN sets new state-of-the-art performance in multiple video recognition benchmarks while achieving faster processing speed.

Se-Ho Kim, Inyong Koo, Inyoung Lee et al.

Denoising diffusion models show remarkable performances in generative tasks, and their potential applications in perception tasks are gaining interest. In this paper, we introduce a novel framework named DiffRef3D which adopts the diffusion process on 3D object detection with point clouds for the first time. Specifically, we formulate the proposal refinement stage of two-stage 3D object detectors as a conditional diffusion process. During training, DiffRef3D gradually adds noise to the residuals between proposals and target objects, then applies the noisy residuals to proposals to generate hypotheses. The refinement module utilizes these hypotheses to denoise the noisy residuals and generate accurate box predictions. In the inference phase, DiffRef3D generates initial hypotheses by sampling noise from a Gaussian distribution as residuals and refines the hypotheses through iterative steps. DiffRef3D is a versatile proposal refinement framework that consistently improves the performance of existing 3D object detection models. We demonstrate the significance of DiffRef3D through extensive experiments on the KITTI benchmark. Code will be available.

Jaehyuk Jang, Wonjun Lee, Kangwook Ko et al.

Prompt tuning has achieved remarkable progress in vision-language models (VLMs) and is recently being adopted for audio-language models (ALMs). However, its generalization ability in ALMs remains largely underexplored. We observe that conventional prompt tuning for ALMs also suffers from the Base-New Tradeoff, and we identify that this issue stems from the disrupted semantic structure of the embedding space. To address this issue, we propose Semantically Expanded Prompt Tuning (SEPT)-a plug-and-play framework that explicitly regularizes the prompt embedding space by incorporating semantic neighbors generated by large language models. SEPT introduces a novel semantic expansion loss with margin constraints that promote intra-class compactness and inter-class separability, thereby enhancing the semantic structure of the prompt embedding space. For comprehensive evaluation, we establish the first benchmark setup for prompt generalization in ALMs, covering both base-to-new generalization and cross-dataset transferability. Extensive experiments demonstrate that SEPT consistently improves generalization performance across multiple prompt tuning baselines, while maintaining computational cost during inference.

Seungju Cho, Hongsin Lee, Changick Kim

Adversarial training is one of the most effective approaches against adversarial attacks. However, adversarial training has primarily been studied in scenarios where data for all classes is provided, with limited research conducted in the context of incremental learning where knowledge is introduced sequentially. In this study, we investigate Adversarially Robust Class Incremental Learning (ARCIL), which deals with adversarial robustness in incremental learning. We first explore a series of baselines that integrate incremental learning with existing adversarial training methods, finding that they lead to conflicts between acquiring new knowledge and retaining past knowledge. Furthermore, we discover that training new knowledge causes the disappearance of a key characteristic in robust models: a flat loss landscape in input space. To address such issues, we propose a novel and robust baseline for ARCIL, named \textbf{FL}atness-preserving \textbf{A}dversarial \textbf{I}ncremental learning for \textbf{R}obustness (\textbf{FLAIR}). Experimental results demonstrate that FLAIR significantly outperforms other baselines. To the best of our knowledge, we are the first to comprehensively investigate the baselines, challenges, and solutions for ARCIL, which we believe represents a significant advance toward achieving real-world robustness. Codes are available at \url{https://github.com/HongsinLee/FLAIR}.

Sumin Lee, Sangmin Woo, Muhammad Adi Nugroho et al.

Due to the distinctive characteristics of sensors, each modality exhibits unique physical properties. For this reason, in the context of multi-modal action recognition, it is important to consider not only the overall action content but also the complementary nature of different modalities. In this paper, we propose a novel network, named Modality Mixer (M-Mixer) network, which effectively leverages and incorporates the complementary information across modalities with the temporal context of actions for action recognition. A key component of our proposed M-Mixer is the Multi-modal Contextualization Unit (MCU), a simple yet effective recurrent unit. Our MCU is responsible for temporally encoding a sequence of one modality (e.g., RGB) with action content features of other modalities (e.g., depth and infrared modalities). This process encourages M-Mixer network to exploit global action content and also to supplement complementary information of other modalities. Furthermore, to extract appropriate complementary information regarding to the given modality settings, we introduce a new module, named Complementary Feature Extraction Module (CFEM). CFEM incorporates sepearte learnable query embeddings for each modality, which guide CFEM to extract complementary information and global action content from the other modalities. As a result, our proposed method outperforms state-of-the-art methods on NTU RGB+D 60, NTU RGB+D 120, and NW-UCLA datasets. Moreover, through comprehensive ablation studies, we further validate the effectiveness of our proposed method.

Minsuk Jang, Yujin Yang, Heeseon Kim et al.

Modern AI-generated videos are photorealistic at the single-frame level, leaving inter-frame dynamics as the main remaining axis for detection. Existing detectors typically handle this temporal evidence in three ways: feeding the full frame sequence to a generic temporal backbone, reducing one dominant temporal cue to fixed video-level descriptors, or comparing temporal features to real-video statistics through a detection metric. These strategies degrade sharply under cross-generator evaluation, where artifact type and timescale vary across generators. On caption-paired benchmark, GenVidBench, we identify two signatures that prior detectors do not jointly exploit: AI-generated videos exhibit smoother frame-to-frame temporal residuals at the pixel level, and more compact trajectories in the semantic feature space, indicating a temporal smoothness gap at both levels. We further observe that, when raw video is fed into a Spiking Neural Networks (SNNs), fake clips elicit firing predominantly at object and motion boundaries, unlike real clips, suggesting that the SNN responds to temporal artifacts localized at edges. These cues are sparse, asynchronous, and concentrated at moments of change, which makes SNNs a natural choice for this task: their event-driven, sparsely-activated dynamics align with the structure of the residual signal in a way that dense ANN backbones do not. Building on this observation, we propose MAST, a detector that processes multi-channel temporal residuals with a spike-driven temporal branch alongside a frozen semantic encoder for cross-generator generalization. On the GenVideo benchmark, MAST achieves 93.14\% mean accuracy across 10 unseen generators under strict cross-generator evaluation, matching or surpassing the strongest ANN-based detectors and demonstrating the practical applicability of SNNs to AI-generated video detection.

Chaehee Song, Minseok Seo, Yeeun Seong et al.

Large language models (LLMs) are typically deployed with fixed parameters, and their performance is often improved by allocating more computation at inference time. While such test-time scaling can be effective, it cannot correct model misconceptions or adapt the model to the specific structure of an individual query. Test-time optimization addresses this limitation by enabling parameter updates during inference, but existing approaches either rely on external data or optimize generic self-supervised objectives that lack query-specific alignment. In this work, we propose Query-Conditioned Test-Time Self-Training (QueST), a framework that adapts model parameters during inference using supervision derived directly from the input query. Our key insight is that the input query itself encodes latent signals sufficient for constructing structurally related problem--solution pairs. Based on this, QueST generates such query-conditioned pairs and uses them as supervision for parameter-efficient fine-tuning at test time. The adapted model is then used to produce the final answer, enabling query-specific adaptation without any external data. Across seven mathematical reasoning benchmarks and the GPQA-Diamond scientific reasoning benchmark, QueST consistently outperforms strong test-time optimization baselines. These results demonstrate that query-conditioned self-training is an effective and practical paradigm for test-time adaptation in LLMs.

Youngjun Kwak, Seonghun Jeong, Yunseung Lee et al.

The budgeted model training challenge aims to train an efficient classification model under resource limitations. To tackle this task in ImageNet-100, we describe a simple yet effective resource-aware backbone search framework composed of profile and instantiation phases. In addition, we employ multi-resolution ensembles to boost inference accuracy on limited resources. The profile phase obeys time and memory constraints to determine the models' optimal batch-size, max epochs, and automatic mixed precision (AMP). And the instantiation phase trains models with the determined parameters from the profile phase. For improving intra-domain generalizations, the multi-resolution ensembles are formed by two-resolution images with randomly applied flips. We present a comprehensive analysis with expensive experiments. Based on our approach, we win first place in International Conference on Computer Vision (ICCV) 2023 Workshop Challenge Track 1 on Resource Efficient Deep Learning for Computer Vision (RCV).

Doyi Kim, Jiseok Ahn, Haewon McJeon et al.

As AI capabilities and deployment accelerate toward a post-AGI era, concerns are growing about electricity demand and carbon emissions from AI computing, yet it is rarely represented explicitly in long term energy-economy-climate scenario models. In such a setting, digital infrastructure scaling may be constrained by power system dynamics. We introduce an AI computing sector into the Global Change Analysis Model (GCAM) and run U.S. scenarios that couple AI service growth with time varying compute energy intensity and economic drivers. We find that service growth does not translate linearly into electricity demand: outcomes depend on efficiency trajectories and demand responsiveness. With sustained efficiency improvements, AI electricity demand remains moderated; with slower or saturating gains, income-driven demand dominates by mid-century. Sensitivity analyses show weak responsiveness to price signals but strong dependence on income growth, implying limited leverage from price-based mechanisms alone. Rather than offering a single forecast, we map conditions under which efficiency-dominant versus demand-dominant regimes emerge, providing a compact template for long run AI electricity-demand scenarios and their implications for power sector emissions.

Minseok Seo, Wonjun Lee, Jaehyuk Jang et al.

Zero-shot depth completion has gained attention for its ability to generalize across environments without sensor-specific datasets or retraining. However, most existing approaches rely on diffusion-based test-time optimization, which is computationally expensive due to iterative denoising. Recent visual-prompt-based methods reduce training cost but still require repeated forward--backward passes through the full frozen network to optimize input-level prompts, resulting in slow inference. In this work, we show that adapting only the decoder is sufficient for effective test-time optimization, as depth foundation models concentrate depth-relevant information within a low-dimensional decoder subspace. Based on this insight, we propose a lightweight test-time adaptation method that updates only this low-dimensional subspace using sparse depth supervision. Our approach achieves state-of-the-art performance, establishing a new Pareto frontier between accuracy and efficiency for test-time adaptation. Extensive experiments on five indoor and outdoor datasets demonstrate consistent improvements over prior methods, highlighting the practicality of fast zero-shot depth completion.

Junyoung Byun, Myung-Joon Kwon, Seungju Cho et al.

Deep neural networks are widely known to be susceptible to adversarial examples, which can cause incorrect predictions through subtle input modifications. These adversarial examples tend to be transferable between models, but targeted attacks still have lower attack success rates due to significant variations in decision boundaries. To enhance the transferability of targeted adversarial examples, we propose introducing competition into the optimization process. Our idea is to craft adversarial perturbations in the presence of two new types of competitor noises: adversarial perturbations towards different target classes and friendly perturbations towards the correct class. With these competitors, even if an adversarial example deceives a network to extract specific features leading to the target class, this disturbance can be suppressed by other competitors. Therefore, within this competition, adversarial examples should take different attack strategies by leveraging more diverse features to overwhelm their interference, leading to improving their transferability to different models. Considering the computational complexity, we efficiently simulate various interference from these two types of competitors in feature space by randomly mixing up stored clean features in the model inference and named this method Clean Feature Mixup (CFM). Our extensive experimental results on the ImageNet-Compatible and CIFAR-10 datasets show that the proposed method outperforms the existing baselines with a clear margin. Our code is available at https://github.com/dreamflake/CFM.

Sangmin Woo, Jinyoung Park, Inyong Koo et al.

Natural Language Video Grounding (NLVG) aims to localize time segments in an untrimmed video according to sentence queries. In this work, we present a new paradigm named Explore-And-Match for NLVG that seamlessly unifies the strengths of two streams of NLVG methods: proposal-free and proposal-based; the former explores the search space to find time segments directly, and the latter matches the predefined time segments with ground truths. To achieve this, we formulate NLVG as a set prediction problem and design an end-to-end trainable Language Video Transformer (LVTR) that can enjoy two favorable properties, which are rich contextualization power and parallel decoding. We train LVTR with two losses. First, temporal localization loss allows time segments of all queries to regress targets (explore). Second, set guidance loss couples every query with their respective target (match). To our surprise, we found that training schedule shows divide-and-conquer-like pattern: time segments are first diversified regardless of the target, then coupled with each target, and fine-tuned to the target again. Moreover, LVTR is highly efficient and effective: it infers faster than previous baselines (by 2X or more) and sets competitive results on two NLVG benchmarks (ActivityCaptions and Charades-STA). Codes are available at https://github.com/sangminwoo/Explore-And-Match.

Seokeon Choi, Taekyung Kim, Minki Jeong et al.

Although supervised person re-identification (Re-ID) methods have shown impressive performance, they suffer from a poor generalization capability on unseen domains. Therefore, generalizable Re-ID has recently attracted growing attention. Many existing methods have employed an instance normalization technique to reduce style variations, but the loss of discriminative information could not be avoided. In this paper, we propose a novel generalizable Re-ID framework, named Meta Batch-Instance Normalization (MetaBIN). Our main idea is to generalize normalization layers by simulating unsuccessful generalization scenarios beforehand in the meta-learning pipeline. To this end, we combine learnable batch-instance normalization layers with meta-learning and investigate the challenging cases caused by both batch and instance normalization layers. Moreover, we diversify the virtual simulations via our meta-train loss accompanied by a cyclic inner-updating manner to boost generalization capability. After all, the MetaBIN framework prevents our model from overfitting to the given source styles and improves the generalization capability to unseen domains without additional data augmentation or complicated network design. Extensive experimental results show that our model outperforms the state-of-the-art methods on the large-scale domain generalization Re-ID benchmark and the cross-domain Re-ID problem. The source code is available at: https://github.com/bismex/MetaBIN.

Seokeon Choi, Sumin Lee, Youngeun Kim et al.

Visible-infrared person re-identification (VI-ReID) is an important task in night-time surveillance applications, since visible cameras are difficult to capture valid appearance information under poor illumination conditions. Compared to traditional person re-identification that handles only the intra-modality discrepancy, VI-ReID suffers from additional cross-modality discrepancy caused by different types of imaging systems. To reduce both intra- and cross-modality discrepancies, we propose a Hierarchical Cross-Modality Disentanglement (Hi-CMD) method, which automatically disentangles ID-discriminative factors and ID-excluded factors from visible-thermal images. We only use ID-discriminative factors for robust cross-modality matching without ID-excluded factors such as pose or illumination. To implement our approach, we introduce an ID-preserving person image generation network and a hierarchical feature learning module. Our generation network learns the disentangled representation by generating a new cross-modality image with different poses and illuminations while preserving a person's identity. At the same time, the feature learning module enables our model to explicitly extract the common ID-discriminative characteristic between visible-infrared images. Extensive experimental results demonstrate that our method outperforms the state-of-the-art methods on two VI-ReID datasets. The source code is available at: https://github.com/bismex/HiCMD.

Sumin Lee, Sungchan Oh, Chanho Jung et al.

Detecting fashion landmarks is a fundamental technique for visual clothing analysis. Due to the large variation and non-rigid deformation of clothes, localizing fashion landmarks suffers from large spatial variances across poses, scales, and styles. Therefore, understanding contextual knowledge of clothes is required for accurate landmark detection. To that end, in this paper, we propose a fashion landmark detection network with a global-local embedding module. The global-local embedding module is based on a non-local operation for capturing long-range dependencies and a subsequent convolution operation for adopting local neighborhood relations. With this processing, the network can consider both global and local contextual knowledge for a clothing image. We demonstrate that our proposed method has an excellent ability to learn advanced deep feature representations for fashion landmark detection. Experimental results on two benchmark datasets show that the proposed network outperforms the state-of-the-art methods. Our code is available at https://github.com/shumming/GLE_FLD.

Minsuk Jang, Hyeonseo Jeong, Minseok Son et al.

While context-based detectors have achieved strong generalization for AI-generated text by measuring distributional inconsistencies, image-based detectors still struggle with overfitting to generator-specific artifacts. We introduce CINEMAE, a novel paradigm for AIGC image detection that adapts the core principles of text detection methods to the visual domain. Our key insight is that Masked AutoEncoder (MAE), trained to reconstruct masked patches conditioned on visible context, naturally encodes semantic consistency expectations. We formalize this reconstruction process probabilistically, computing conditional Negative Log-Likelihood (NLL, p(masked | visible)) to quantify local semantic anomalies. By aggregating these patch-level statistics with global MAE features through learned fusion, CINEMAE achieves strong cross-generator generalization. Trained exclusively on Stable Diffusion v1.4, our method achieves over 95% accuracy on all eight unseen generators in the GenImage benchmark, substantially outperforming state-of-the-art detectors. This demonstrates that context-conditional reconstruction uncertainty provides a robust, transferable signal for AIGC detection.

Sangmin Woo, Byeongjun Park, Hyojun Go et al.

Recent progress in single-image 3D generation highlights the importance of multi-view coherency, leveraging 3D priors from large-scale diffusion models pretrained on Internet-scale images. However, the aspect of novel-view diversity remains underexplored within the research landscape due to the ambiguity in converting a 2D image into 3D content, where numerous potential shapes can emerge. Here, we aim to address this research gap by simultaneously addressing both consistency and diversity. Yet, striking a balance between these two aspects poses a considerable challenge due to their inherent trade-offs. This work introduces HarmonyView, a simple yet effective diffusion sampling technique adept at decomposing two intricate aspects in single-image 3D generation: consistency and diversity. This approach paves the way for a more nuanced exploration of the two critical dimensions within the sampling process. Moreover, we propose a new evaluation metric based on CLIP image and text encoders to comprehensively assess the diversity of the generated views, which closely aligns with human evaluators' judgments. In experiments, HarmonyView achieves a harmonious balance, demonstrating a win-win scenario in both consistency and diversity.

Jaehyuk Jang, Yooseung Wang, Changick Kim

Recently, multimodal prompting, which introduces learnable missing-aware prompts for all missing modality cases, has exhibited impressive performance. However, it encounters two critical issues: 1) The number of prompts grows exponentially as the number of modalities increases; and 2) It lacks robustness in scenarios with different missing modality settings between training and inference. In this paper, we propose a simple yet effective prompt design to address these challenges. Instead of using missing-aware prompts, we utilize prompts as modality-specific tokens, enabling them to capture the unique characteristics of each modality. Furthermore, our prompt design leverages orthogonality between prompts as a key element to learn distinct information across different modalities and promote diversity in the learned representations. Extensive experiments demonstrate that our prompt design enhances both performance and robustness while reducing the number of prompts.

Myung-Joon Kwon, Wonjun Lee, Seung-Hun Nam et al.

Most techniques approach the problem of image forgery localization as a binary segmentation task, training neural networks to label original areas as 0 and forged areas as 1. In contrast, we tackle this issue from a more fundamental perspective by partitioning images according to their originating sources. To this end, we propose Segment Any Forged Image Region (SAFIRE), which solves forgery localization using point prompting. Each point on an image is used to segment the source region containing itself. This allows us to partition images into multiple source regions, a capability achieved for the first time. Additionally, rather than memorizing certain forgery traces, SAFIRE naturally focuses on uniform characteristics within each source region. This approach leads to more stable and effective learning, achieving superior performance in both the new task and the traditional binary forgery localization.

Hyojun Go, Byeongjun Park, Jiho Jang et al.

Text-based generation and editing of 3D scenes hold significant potential for streamlining content creation through intuitive user interactions. While recent advances leverage 3D Gaussian Splatting (3DGS) for high-fidelity and real-time rendering, existing methods are often specialized and task-focused, lacking a unified framework for both generation and editing. In this paper, we introduce SplatFlow, a comprehensive framework that addresses this gap by enabling direct 3DGS generation and editing. SplatFlow comprises two main components: a multi-view rectified flow (RF) model and a Gaussian Splatting Decoder (GSDecoder). The multi-view RF model operates in latent space, generating multi-view images, depths, and camera poses simultaneously, conditioned on text prompts, thus addressing challenges like diverse scene scales and complex camera trajectories in real-world settings. Then, the GSDecoder efficiently translates these latent outputs into 3DGS representations through a feed-forward 3DGS method. Leveraging training-free inversion and inpainting techniques, SplatFlow enables seamless 3DGS editing and supports a broad range of 3D tasks-including object editing, novel view synthesis, and camera pose estimation-within a unified framework without requiring additional complex pipelines. We validate SplatFlow's capabilities on the MVImgNet and DL3DV-7K datasets, demonstrating its versatility and effectiveness in various 3D generation, editing, and inpainting-based tasks.

Hee-Seon Kim, Minbeom Kim, Changick Kim

Large Vision-Language Models (VLMs) have demonstrated remarkable performance across multimodal tasks by integrating vision encoders with large language models (LLMs). However, these models remain vulnerable to adversarial attacks. Among such attacks, Universal Adversarial Perturbations (UAPs) are especially powerful, as a single optimized perturbation can mislead the model across various input images. In this work, we introduce a novel UAP specifically designed for VLMs: the Doubly-Universal Adversarial Perturbation (Doubly-UAP), capable of universally deceiving VLMs across both image and text inputs. To successfully disrupt the vision encoder's fundamental process, we analyze the core components of the attention mechanism. After identifying value vectors in the middle-to-late layers as the most vulnerable, we optimize Doubly-UAP in a label-free manner with a frozen model. Despite being developed as a black-box to the LLM, Doubly-UAP achieves high attack success rates on VLMs, consistently outperforming baseline methods across vision-language tasks. Extensive ablation studies and analyses further demonstrate the robustness of Doubly-UAP and provide insights into how it influences internal attention mechanisms.

Hyojun Go, Byeongjun Park, Hyelin Nam et al.

We propose VideoRFSplat, a direct text-to-3D model leveraging a video generation model to generate realistic 3D Gaussian Splatting (3DGS) for unbounded real-world scenes. To generate diverse camera poses and unbounded spatial extent of real-world scenes, while ensuring generalization to arbitrary text prompts, previous methods fine-tune 2D generative models to jointly model camera poses and multi-view images. However, these methods suffer from instability when extending 2D generative models to joint modeling due to the modality gap, which necessitates additional models to stabilize training and inference. In this work, we propose an architecture and a sampling strategy to jointly model multi-view images and camera poses when fine-tuning a video generation model. Our core idea is a dual-stream architecture that attaches a dedicated pose generation model alongside a pre-trained video generation model via communication blocks, generating multi-view images and camera poses through separate streams. This design reduces interference between the pose and image modalities. Additionally, we propose an asynchronous sampling strategy that denoises camera poses faster than multi-view images, allowing rapidly denoised poses to condition multi-view generation, reducing mutual ambiguity and enhancing cross-modal consistency. Trained on multiple large-scale real-world datasets (RealEstate10K, MVImgNet, DL3DV-10K, ACID), VideoRFSplat outperforms existing text-to-3D direct generation methods that heavily depend on post-hoc refinement via score distillation sampling, achieving superior results without such refinement.

Byeongjun Park, Hyojun Go, Hyelin Nam et al.

Recent progress in 3D/4D scene generation emphasizes the importance of physical alignment throughout video generation and scene reconstruction. However, existing methods improve the alignment separately at each stage, making it difficult to manage subtle misalignments arising from another stage. Here, we present SteerX, a zero-shot inference-time steering method that unifies scene reconstruction into the generation process, tilting data distributions toward better geometric alignment. To this end, we introduce two geometric reward functions for 3D/4D scene generation by using pose-free feed-forward scene reconstruction models. Through extensive experiments, we demonstrate the effectiveness of SteerX in improving 3D/4D scene generation.

Minseok Son, Inyong Koo, Jinyoung Park et al.

When trained with severely imbalanced data, deep neural networks often struggle to accurately recognize classes with only a few samples. Previous studies in long-tailed recognition have attempted to rebalance biased learning using known sample distributions, primarily addressing different classification difficulties at the class level. However, these approaches often overlook the instance difficulty variation within each class. In this paper, we propose a difficulty-aware balancing margin (DBM) loss, which considers both class imbalance and instance difficulty. DBM loss comprises two components: a class-wise margin to mitigate learning bias caused by imbalanced class frequencies, and an instance-wise margin assigned to hard positive samples based on their individual difficulty. DBM loss improves class discriminativity by assigning larger margins to more difficult samples. Our method seamlessly combines with existing approaches and consistently improves performance across various long-tailed recognition benchmarks.

Seokil Ham, Hee-Seon Kim, Sangmin Woo et al.

Despite the growing interest in Mamba architecture as a potential replacement for Transformer architecture, parameter-efficient fine-tuning (PEFT) approaches for Mamba remain largely unexplored. In our study, we introduce two key insights-driven strategies for PEFT in Mamba architecture: (1) While state-space models (SSMs) have been regarded as the cornerstone of Mamba architecture, then expected to play a primary role in transfer learning, our findings reveal that Projectors -- not SSMs -- are the predominant contributors to transfer learning. (2) Based on our observation, we propose a novel PEFT method specialized to Mamba architecture: Projector-targeted Diagonal-centric Linear Transformation (ProDiaL). ProDiaL focuses on optimizing only the pretrained Projectors for new tasks through diagonal-centric linear transformation matrices, without directly fine-tuning the Projector weights. This targeted approach allows efficient task adaptation, utilizing less than 1% of the total parameters, and exhibits strong performance across both vision and language Mamba models, highlighting its versatility and effectiveness.

Sumin Lee, Yooseung Wang, Sangmin Woo et al.

Panoramic Activity Recognition (PAR) seeks to identify diverse human activities across different scales, from individual actions to social group and global activities in crowded panoramic scenes. PAR presents two major challenges: 1) recognizing the nuanced interactions among numerous individuals and 2) understanding multi-granular human activities. To address these, we propose Social Proximity-aware Dual-Path Network (SPDP-Net) based on two key design principles. First, while previous works often focus on spatial distance among individuals within an image, we argue to consider the spatio-temporal proximity. It is crucial for individual relation encoding to correctly understand social dynamics. Secondly, deviating from existing hierarchical approaches (individual-to-social-to-global activity), we introduce a dual-path architecture for multi-granular activity recognition. This architecture comprises individual-to-global and individual-to-social paths, mutually reinforcing each other's task with global-local context through multiple layers. Through extensive experiments, we validate the effectiveness of the spatio-temporal proximity among individuals and the dual-path architecture in PAR. Furthermore, SPDP-Net achieves new state-of-the-art performance with 46.5\% of overall F1 score on JRDB-PAR dataset.

Hongsin Lee, Seungju Cho, Changick Kim

Adversarial training significantly improves adversarial robustness, but superior performance is primarily attained with large models. This substantial performance gap for smaller models has spurred active research into adversarial distillation (AD) to mitigate the difference. Existing AD methods leverage the teacher's logits as a guide. In contrast to these approaches, we aim to transfer another piece of knowledge from the teacher, the input gradient. In this paper, we propose a distillation module termed Indirect Gradient Distillation Module (IGDM) that indirectly matches the student's input gradient with that of the teacher. Experimental results show that IGDM seamlessly integrates with existing AD methods, significantly enhancing their performance. Particularly, utilizing IGDM on the CIFAR-100 dataset improves the AutoAttack accuracy from 28.06% to 30.32% with the ResNet-18 architecture and from 26.18% to 29.32% with the MobileNetV2 architecture when integrated into the SOTA method without additional data augmentation.

Younhun Kim, Myung-Joon Kwon, Wonjun Lee et al.

Previous Deepfake detection methods perform well within their training domains, but their effectiveness diminishes significantly with new synthesis techniques. Recent studies have revealed that detection models often create decision boundaries based on facial identity rather than synthetic artifacts, resulting in poor performance on cross-domain datasets. To address this limitation, we propose Facial Recognition Identity Attenuation (FRIDAY), a novel training method that mitigates facial identity influence using a face recognizer. Specifically, we first train a face recognizer using the same backbone as the Deepfake detector. The recognizer is then frozen and employed during the detector's training to reduce facial identity information. This is achieved by feeding input images into both the recognizer and the detector, and minimizing the similarity of their feature embeddings through our Facial Identity Attenuating loss. This process encourages the detector to generate embeddings distinct from the recognizer, effectively reducing the impact of facial identity. Extensive experiments demonstrate that our approach significantly enhances detection performance on both in-domain and cross-domain datasets.

Minseok Seo, Mark Hamilton, Changick Kim

We present \textbf{Upsample Anything}, a lightweight test-time optimization (TTO) framework that restores low-resolution features to high-resolution, pixel-wise outputs without any training. Although Vision Foundation Models demonstrate strong generalization across diverse downstream tasks, their representations are typically downsampled by 14x/16x (e.g., ViT), which limits their direct use in pixel-level applications. Existing feature upsampling approaches depend on dataset-specific retraining or heavy implicit optimization, restricting scalability and generalization. Upsample Anything addresses these issues through a simple per-image optimization that learns an anisotropic Gaussian kernel combining spatial and range cues, effectively bridging Gaussian Splatting and Joint Bilateral Upsampling. The learned kernel acts as a universal, edge-aware operator that transfers seamlessly across architectures and modalities, enabling precise high-resolution reconstruction of features, depth, or probability maps. It runs in only $\approx0.419 \text{s}$ per 224x224 image and achieves state-of-the-art performance on semantic segmentation, depth estimation, and both depth and probability map upsampling.

Youngju Yoo, Seho Kim, Changick Kim

3D instance segmentation is crucial for understanding complex 3D environments, yet fully supervised methods require dense point-level annotations, resulting in substantial annotation costs and labor overhead. To mitigate this, box-level annotations have been explored as a weaker but more scalable form of supervision. However, box annotations inherently introduce ambiguity in overlapping regions, making accurate point-to-instance assignment challenging. Recent methods address this ambiguity by generating pseudo-masks through training a dedicated pseudo-labeler in an additional training stage. However, such two-stage pipelines often increase overall training time and complexity, hinder end-to-end optimization. To overcome these challenges, we propose BEEP3D-Box-supervised End-to-End Pseudo-mask generation for 3D instance segmentation. BEEP3D adopts a student-teacher framework, where the teacher model serves as a pseudo-labeler and is updated by the student model via an Exponential Moving Average. To better guide the teacher model to generate precise pseudo-masks, we introduce an instance center-based query refinement that enhances position query localization and leverages features near instance centers. Additionally, we design two novel losses-query consistency loss and masked feature consistency loss-to align semantic and geometric signals between predictions and pseudo-masks. Extensive experiments on ScanNetV2 and S3DIS datasets demonstrate that BEEP3D achieves competitive or superior performance compared to state-of-the-art weakly supervised methods while remaining computationally efficient.

Jinyoung Park, Minseok Son, Changick Kim

State Space Models (SSMs) have emerged as powerful architectures in computer vision, yet improving their computational efficiency remains crucial for practical and scalable deployment.While token reduction serves as an effective approach for model efficiency, applying it to SSMs requires careful consideration of their unique sequential modeling capabilities.In this work, we propose MaMe, a token-merging strategy tailored for SSM-based vision models.MaMe addresses two key challenges: quantifying token importance and preserving sequential properties. Our approach leverages the state transition parameter $\mathbfΔ$ as an informativeness measure and introduces strategic token arrangements to preserve sequential information flow.Extensive experiments demonstrate that MaMe achieves superior efficiency-performance trade-offs for both fine-tuned and off-the-shelf models. Particularly, our approach maintains robustness even under aggressive token reduction where existing methods undergo significant performance degradation.Beyond image classification, MaMe shows strong generalization capabilities across video and audio domains, establishing an effective approach for enhancing efficiency in diverse SSM applications.

Minsuk Jang, Changick Kim

Current Spiking Neural Networks (SNNs) underutilize the temporal dynamics inherent in spike-based processing, relying primarily on rate coding while overlooking precise timing information that provides rich computational cues. We propose SPARTA (Spiking Priority Attention with Resource-Adaptive Temporal Allocation), a framework that leverages heterogeneous neuron dynamics and spike-timing information to enable efficient sparse attention. SPARTA prioritizes tokens based on temporal cues, including firing patterns, spike timing, and inter-spike intervals, achieving 65.4% sparsity through competitive gating. By selecting only the most salient tokens, SPARTA reduces attention complexity from O(N^2) to O(K^2) with k << n, while maintaining high accuracy. Our method achieves state-of-the-art performance on DVS-Gesture (98.78%) and competitive results on CIFAR10-DVS (83.06%) and CIFAR-10 (95.3%), demonstrating that exploiting spike timing dynamics improves both computational efficiency and accuracy.

Younghun Kim, Minsuk Jang, Myung-Joon Kwon et al.

Face identity provides a powerful signal for deepfake detection. Prior studies show that even when not explicitly modeled, classifiers often learn identity features implicitly. This has led to conflicting views: some suppress identity cues to reduce bias, while others rely on them as forensic evidence. To reconcile these views, we analyze two hypotheses: (1) whether face identity alone is discriminative for detecting deepfakes, and (2) whether such identity features generalize poorly across manipulation methods. Our experiments confirm that identity is informative but context-dependent. While some manipulations preserve identity-consistent artifacts, others distort identity cues and harm generalization. We argue that identity features should neither be blindly suppressed nor relied upon, but instead be explicitly modeled and adaptively controlled based on per-sample relevance. We propose \textbf{SELFI} (\textbf{SEL}ective \textbf{F}usion of \textbf{I}dentity), a generalizable detection framework that dynamically modulates identity usage. SELFI consists of: (1) a Forgery-Aware Identity Adapter (FAIA) that extracts identity embeddings from a frozen face recognition model and projects them into a forgery-relevant space via auxiliary supervision; and (2) an Identity-Aware Fusion Module (IAFM) that selectively integrates identity and visual features using a relevance-guided fusion mechanism. Experiments on four benchmarks show that SELFI improves cross-manipulation generalization, outperforming prior methods by an average of 3.1\% AUC. On the challenging DFDC dataset, SELFI exceeds the previous best by 6\%. Code will be released upon paper acceptance.