arXiv subjects

Songkuk Kim

Publications and source records attributed to Songkuk Kim.

At least 19 recordsLinked to original sources

arXiv2026

Aligning Sentence Embeddings to Human Concepts via Sparse Autoencoders

Dense sentence embeddings are fundamental to modern Retrieval-Augmented Generation (RAG) systems but suffer from a lack of interpretability due to feature superposition. This opacity hinders the alignment of retrieval processes with human intent, as the entangled representations are difficult to analyze or control. In this work, we propose a method to disentangle the dense representations of sentence transformers (e.g., E5) into human-interpretable concepts using Top-k Sparse Autoencoders (SAEs). We demonstrate that these disentangled features align with specific semantic, syntactic, and pragmatic categories. Furthermore, we introduce an activation steering mechanism that allows for precise intervention in the retrieval process. By clamping specific latent features, we show that it is possible to re-rank search results to better align with user constraints without retraining the backbone model. Our findings suggest that SAE-based decomposition offers a viable path toward transparent and steerable neural information retrieval.

cs.IR↗

arXiv2026

What Do Students Learn? A Feature-Level Analysis of Dark Knowledge

Knowledge Distillation (KD) is a powerful tool for model compression, yet the precise mechanisms by which student models acquire feature representations remain underexplored. In this work, we analyze student feature learning using the Interaction Tensor framework. Our analysis reveals that effective KD acts as a regularizer that prunes low-frequency, sample-specific features, encouraging the student to rely on a compact set of highly reusable features. Crucially, we observe that the dataset-level confusion matrix contains structural information analogous to the teacher's "Dark Knowledge." Leveraging this insight, we propose Confusion Distillation (CD), a teacher-free self-distillation method that utilizes the model's own evolving )confusion patterns as dynamic soft targets. CD achieves competitive performance on ResNet-34 and ResNet-50 for CIFAR-100, outperforming existing self-distillation methods like CS-KD and PS-KD by 1.2% while offering a computationally efficient alternative to standard KD.

cs.LG↗

arXiv2026

DebFilter: Eradicating Biases Stashed in Value

Text-to-image diffusion models, which are theoretically equivalent to score-based generative models, generate images through a multi-step denoising process guided by text embeddings extracted from pretrained vision-language models such as CLIP. However, these text embeddings inherently encode social and semantic biases -- such as those related to gender and age -- that are subsequently propagated and amplified through the guidance mechanism, along with the model's training on large-scale datasets that are imbalanced with respect to these bias-related concepts, often leading to skewed outputs in text-to-image generation. We propose DebFilter, a lightweight and training-free framework for mitigating such biases in text-to-image diffusion models. Observing that the model's error prediction at each denoising step is primarily influenced by cross-attention dynamics, we introduce a bias-correction strategy that adjusts the value components within cross-attention. Specifically, we apply a fixed offset to the slice of guidance embedding, effectively steering the semantic direction of cross-attention values toward unbiased representations. This adjustment reconfigures the score landscape to produce balanced outputs while maintaining alignment with the intended text semantics. Unlike prior approaches that rely on fine-tuning or retraining, DebFilter operates entirely at inference time, requiring no additional data or model updates. Our results demonstrate that this method effectively mitigates social biases in generated images, offering an efficient and scalable pathway toward fairer and more inclusive text-to-image generation.

cs.CV↗

arXiv2026

Sparsity as a Key: Unlocking New Insights from Latent Structures for Out-of-Distribution Detection

Sparse Autoencoders (SAEs) have demonstrated significant success in interpreting Large Language Models (LLMs) by decomposing dense representations into sparse, semantic components. However, their potential for analyzing Vision Transformers (ViTs) remains largely under-explored. In this work, we present the first application of SAEs to the ViT [CLS] token for out-of-distribution (OOD) detection, addressing the limitation of existing methods that rely on entangled feature representations. We propose a novel framework utilizing a Top-k SAE to disentangle the dense [CLS] features into a structured latent space. Through this analysis, we reveal that in-distribution (ID) data exhibits consistent, class-specific activation patterns, which we formalize as Class Activation Profiles (CAPs). Our study uncovers a key structural invariant: while ID samples preserve a stable pattern within CAPs, OOD samples systematically disrupt this structure. Leveraging this insight, we introduce a scoring function based on the divergence of core energy profiles to quantify the deviation from ideal activation profiles. Our method achieves strong results on the FPR95 metric, critical for safety-sensitive applications across multiple benchmarks, while also achieving competitive AUROC. Overall, our findings demonstrate that the sparse, disentangled features revealed by SAEs can serve as a powerful, interpretable tool for robust OOD detection in vision models.

cs.CV↗

arXiv2026

Temperature Scaling Attack Disrupting Model Confidence in Federated Learning

Predictive confidence serves as a foundational control signal in mission-critical systems, directly governing risk-aware logic such as escalation, abstention, and conservative fallback. While prior federated learning attacks predominantly target accuracy or implant backdoors, we identify confidence calibration as a distinct attack objective. We present the Temperature Scaling Attack (TSA), a training-time attack that degrades calibration while preserving accuracy. By injecting temperature scaling with learning rate-temperature coupling during local training, TSA shifts model confidence while keeping predictive accuracy and common optimization signals close to benign training. We provide a convergence analysis under non-IID settings, showing that the coupling controls the primary update scale while leaving a bounded temperature-induced residual, yielding the standard non-convex FL convergence structure with an additional residual term. Across three benchmarks, TSA substantially shifts calibration (e.g., 145% error increase on CIFAR-100) with <2% accuracy change, and remains effective under robust aggregation and post-hoc calibration defenses. Case studies further show up to a 7.2x increase in missed verifications in healthcare and severe confidence-gating failures in autonomous driving, even when accuracy is unchanged. Overall, our results establish calibration integrity as a critical attack surface in federated learning.

cs.LG↗

arXiv2025

Symbol Distributions in Semantic Communications: A Source-Channel Equilibrium Perspective

Semantic communication systems often use end-to-end neural networks to map input data into continuous symbols. These symbols, which are essentially neural network features, have fixed dimensions and often exhibit heavy-tailed distributions. However, the mechanism behind this distributional shape remains underexplored due to the end-to-end nature of encoder training, hindering systematic analysis and design. In this paper, we propose a parametric model for semantic symbol distributions. We model end-to-end training as inducing two coupled pressures on the symbol distribution: a source pressure that favors power allocation minimizing the average description cost, and a channel pressure that favors distributions with higher channel utilization. Under surrogate objectives that capture these effects, we obtain a Student's t-distribution as a model for the semantic symbols. Experiments on image-based semantic systems show that the model closely predicts how the shape parameter varies with (i) explicit symbol rate control and (ii) dataset entropy variability. Furthermore, enforcing a target symbol distribution via regularization (e.g., a Gaussian prior) improves training convergence, which is consistent with our hypothesis.

cs.IT↗

arXiv2025

How Diffusion Models Memorize

Despite their success in image generation, diffusion models can memorize training data, raising serious privacy and copyright concerns. Although prior work has sought to characterize, detect, and mitigate memorization, the fundamental question of why and how it occurs remains unresolved. In this paper, we revisit the diffusion and denoising process and analyze latent space dynamics to address the question: "How do diffusion models memorize?" We show that memorization is driven by the overestimation of training samples during early denoising, which reduces diversity, collapses denoising trajectories, and accelerates convergence toward the memorized image. Specifically: (i) memorization cannot be explained by overfitting alone, as training loss is larger under memorization due to classifier-free guidance amplifying predictions and inducing overestimation; (ii) memorized prompts inject training images into noise predictions, forcing latent trajectories to converge and steering denoising toward their paired samples; and (iii) a decomposition of intermediate latents reveals how initial randomness is quickly suppressed and replaced by memorized content, with deviations from the theoretical denoising schedule correlating almost perfectly with memorization severity. Together, these results identify early overestimation as the central underlying mechanism of memorization in diffusion models.

cs.CV↗

arXiv2025

Toward Storage-Aware Learning with Compressed Data An Empirical Exploratory Study on JPEG

On-device machine learning is often constrained by limited storage, particularly in continuous data collection scenarios. This paper presents an empirical study on storage-aware learning, focusing on the trade-off between data quantity and quality via compression. We demonstrate that naive strategies, such as uniform data dropping or one-size-fits-all compression, are suboptimal. Our findings further reveal that data samples exhibit varying sensitivities to compression, supporting the feasibility of a sample-wise adaptive compression strategy. These insights provide a foundation for developing a new class of storage-aware learning systems. The primary contribution of this work is the systematic characterization of this under-explored challenge, offering valuable insights that advance the understanding of storage-aware learning.

cs.LG↗

arXiv2025

Bridging Neural Networks and Wireless Systems with MIMO-OFDM Semantic Communications

Semantic communications aim to enhance transmission efficiency by jointly optimizing source coding, channel coding, and modulation. While prior research has demonstrated promising performance in simulations, real-world implementations often face significant challenges, including noise variability and nonlinear distortions, leading to performance gaps. This article investigates these challenges in a multiple-input multiple-output (MIMO) and orthogonal frequency-division multiplexing (OFDM)-based semantic communication system, focusing on the practical impacts of power amplifier (PA) nonlinearity and peak-to-average power ratio (PAPR) variations. Our analysis identifies frequency selectivity of the actual channel as a critical factor in performance degradation and demonstrates that targeted mitigation strategies can enable semantic systems to approach theoretical performance. By addressing key limitations in existing designs, we provide actionable insights for advancing semantic communications in practical wireless environments. This work establishes a foundation for bridging the gap between theoretical models and real-world deployment, highlighting essential considerations for system design and optimization.

cs.IT↗

arXiv2024

Tazza: Shuffling Neural Network Parameters for Secure and Private Federated Learning

Federated learning enables decentralized model training without sharing raw data, preserving data privacy. However, its vulnerability towards critical security threats, such as gradient inversion and model poisoning by malicious clients, remain unresolved. Existing solutions often address these issues separately, sacrificing either system robustness or model accuracy. This work introduces Tazza, a secure and efficient federated learning framework that simultaneously addresses both challenges. By leveraging the permutation equivariance and invariance properties of neural networks via weight shuffling and shuffled model validation, Tazza enhances resilience against diverse poisoning attacks, while ensuring data confidentiality and high model accuracy. Comprehensive evaluations on various datasets and embedded platforms show that Tazza achieves robust defense with up to 6.7x improved computational efficiency compared to alternative schemes, without compromising performance.

cs.LG↗

arXiv2024

DeTrigger: A Gradient-Centric Approach to Backdoor Attack Mitigation in Federated Learning

Federated Learning (FL) enables collaborative model training across distributed devices while preserving local data privacy, making it ideal for mobile and embedded systems. However, the decentralized nature of FL also opens vulnerabilities to model poisoning attacks, particularly backdoor attacks, where adversaries implant trigger patterns to manipulate model predictions. In this paper, we propose DeTrigger, a scalable and efficient backdoor-robust federated learning framework that leverages insights from adversarial attack methodologies. By employing gradient analysis with temperature scaling, DeTrigger detects and isolates backdoor triggers, allowing for precise model weight pruning of backdoor activations without sacrificing benign model knowledge. Extensive evaluations across four widely used datasets demonstrate that DeTrigger achieves up to 251x faster detection than traditional methods and mitigates backdoor attacks by up to 98.9%, with minimal impact on global model accuracy. Our findings establish DeTrigger as a robust and scalable solution to protect federated learning environments against sophisticated backdoor threats.

cs.LG↗

arXiv2024

Improving Local Training in Federated Learning via Temperature Scaling

Federated learning is inherently hampered by data heterogeneity: non-i.i.d. training data over local clients. We propose a novel model training approach for federated learning, FLex&Chill, which exploits the Logit Chilling method. Through extensive evaluations, we demonstrate that, in the presence of non-i.i.d. data characteristics inherent in federated learning systems, this approach can expedite model convergence and improve inference accuracy. Quantitatively, from our experiments, we observe up to 6X improvement in the global federated learning model convergence time, and up to 3.37% improvement in inference accuracy.

cs.LG↗

arXiv2023

Curved Representation Space of Vision Transformers

Neural networks with self-attention (a.k.a. Transformers) like ViT and Swin have emerged as a better alternative to traditional convolutional neural networks (CNNs). However, our understanding of how the new architecture works is still limited. In this paper, we focus on the phenomenon that Transformers show higher robustness against corruptions than CNNs, while not being overconfident. This is contrary to the intuition that robustness increases with confidence. We resolve this contradiction by empirically investigating how the output of the penultimate layer moves in the representation space as the input data moves linearly within a small area. In particular, we show the following. (1) While CNNs exhibit fairly linear relationship between the input and output movements, Transformers show nonlinear relationship for some data. For those data, the output of Transformers moves in a curved trajectory as the input moves linearly. (2) When a data is located in a curved region, it is hard to move it out of the decision region since the output moves along a curved trajectory instead of a straight line to the decision boundary, resulting in high robustness of Transformers. (3) If a data is slightly modified to jump out of the curved region, the movements afterwards become linear and the output goes to the decision boundary directly. In other words, there does exist a decision boundary near the data, which is hard to find only because of the curved representation space. This explains the underconfident prediction of Transformers. Also, we examine mathematical properties of the attention operation that induce nonlinear response to linear perturbation. Finally, we share our additional findings, regarding what contributes to the curved representation space of Transformers, and how the curvedness evolves during training.

cs.CV↗

arXiv2023

On the Role of ViT and CNN in Semantic Communications: Analysis and Prototype Validation

Semantic communications have shown promising advancements by optimizing source and channel coding jointly. However, the dynamics of these systems remain understudied, limiting research and performance gains. Inspired by the robustness of Vision Transformers (ViTs) in handling image nuisances, we propose a ViT-based model for semantic communications. Our approach achieves a peak signal-to-noise ratio (PSNR) gain of +0.5 dB over convolutional neural network variants. We introduce novel measures, average cosine similarity and Fourier analysis, to analyze the inner workings of semantic communications and optimize the system's performance. We also validate our approach through a real wireless channel prototype using software-defined radio (SDR). To the best of our knowledge, this is the first investigation of the fundamental workings of a semantic communications system, accompanied by the pioneering hardware implementation. To facilitate reproducibility and encourage further research, we provide open-source code, including neural network implementations and LabVIEW codes for SDR-based wireless transmission systems.

eess.SP↗

arXiv2022

Blurs Behave Like Ensembles: Spatial Smoothings to Improve Accuracy, Uncertainty, and Robustness

Neural network ensembles, such as Bayesian neural networks (BNNs), have shown success in the areas of uncertainty estimation and robustness. However, a crucial challenge prohibits their use in practice. BNNs require a large number of predictions to produce reliable results, leading to a significant increase in computational cost. To alleviate this issue, we propose spatial smoothing, a method that spatially ensembles neighboring feature map points of convolutional neural networks. By simply adding a few blur layers to the models, we empirically show that spatial smoothing improves accuracy, uncertainty estimation, and robustness of BNNs across a whole range of ensemble sizes. In particular, BNNs incorporating spatial smoothing achieve high predictive performance merely with a handful of ensembles. Moreover, this method also can be applied to canonical deterministic neural networks to improve the performances. A number of evidences suggest that the improvements can be attributed to the stabilized feature maps and the smoothing of the loss landscape. In addition, we provide a fundamental explanation for prior works - namely, global average pooling, pre-activation, and ReLU6 - by addressing them as special cases of spatial smoothing. These not only enhance accuracy, but also improve uncertainty estimation and robustness by making the loss landscape smoother in the same manner as spatial smoothing. The code is available at https://github.com/xxxnell/spatial-smoothing.

cs.LG↗

arXiv2022

How Do Vision Transformers Work?

The success of multi-head self-attentions (MSAs) for computer vision is now indisputable. However, little is known about how MSAs work. We present fundamental explanations to help better understand the nature of MSAs. In particular, we demonstrate the following properties of MSAs and Vision Transformers (ViTs): (1) MSAs improve not only accuracy but also generalization by flattening the loss landscapes. Such improvement is primarily attributable to their data specificity, not long-range dependency. On the other hand, ViTs suffer from non-convex losses. Large datasets and loss landscape smoothing methods alleviate this problem; (2) MSAs and Convs exhibit opposite behaviors. For example, MSAs are low-pass filters, but Convs are high-pass filters. Therefore, MSAs and Convs are complementary; (3) Multi-stage neural networks behave like a series connection of small individual models. In addition, MSAs at the end of a stage play a key role in prediction. Based on these insights, we propose AlterNet, a model in which Conv blocks at the end of a stage are replaced with MSA blocks. AlterNet outperforms CNNs not only in large data regimes but also in small data regimes. The code is available at https://github.com/xxxnell/how-do-vits-work.

cs.CV↗

arXiv2022

Demo: Real-Time Semantic Communications with a Vision Transformer

Semantic communications are expected to enable the more effective delivery of meaning rather than a precise transfer of symbols. In this paper, we propose an end-to-end deep neural network-based architecture for image transmission and demonstrate its feasibility in a real-time wireless channel by implementing a prototype based on a field-programmable gate array (FPGA). We demonstrate that this system outperforms the traditional 256-quadrature amplitude modulation system in the low signal-to-noise ratio regime with the popular CIFAR-10 dataset. To the best of our knowledge, this is the first work that implements and investigates real-time semantic communications with a vision transformer.

eess.SP↗

arXiv2021

Vector Quantized Bayesian Neural Network Inference for Data Streams

Bayesian neural networks (BNN) can estimate the uncertainty in predictions, as opposed to non-Bayesian neural networks (NNs). However, BNNs have been far less widely used than non-Bayesian NNs in practice since they need iterative NN executions to predict a result for one data, and it gives rise to prohibitive computational cost. This computational burden is a critical problem when processing data streams with low-latency. To address this problem, we propose a novel model VQ-BNN, which approximates BNN inference for data streams. In order to reduce the computational burden, VQ-BNN inference predicts NN only once and compensates the result with previously memorized predictions. To be specific, VQ-BNN inference for data streams is given by temporal exponential smoothing of recent predictions. The computational cost of this model is almost the same as that of non-Bayesian NNs. Experiments including semantic segmentation on real-world data show that this model performs significantly faster than BNNs while estimating predictive results comparable to or superior to the results of BNNs.

cs.LG↗