Sooel Son

Minseon Kim, Hyeonjeong Ha, Sooel Son et al.

Recently, unsupervised adversarial training (AT) has been highlighted as a means of achieving robustness in models without any label information. Previous studies in unsupervised AT have mostly focused on implementing self-supervised learning (SSL) frameworks, which maximize the instance-wise classification loss to generate adversarial examples. However, we observe that simply maximizing the self-supervised training loss with an untargeted adversarial attack often results in generating ineffective adversaries that may not help improve the robustness of the trained model, especially for non-contrastive SSL frameworks without negative examples. To tackle this problem, we propose a novel positive mining for targeted adversarial attack to generate effective adversaries for adversarial SSL frameworks. Specifically, we introduce an algorithm that selects the most confusing yet similar target example for a given instance based on entropy and similarity, and subsequently perturbs the given instance towards the selected target. Our method demonstrates significant enhancements in robustness when applied to non-contrastive SSL frameworks, and less but consistent robustness improvements with contrastive SSL frameworks, on the benchmark datasets.

Joonsung Jeon, Woo Jae Kim, Suhyeon Ha et al.

Latent diffusion models have achieved remarkable success in high-fidelity text-to-image generation, but their tendency to memorize training data raises critical privacy and intellectual property concerns. Membership inference attacks (MIAs) provide a principled way to audit such memorization by determining whether a given sample was included in training. However, existing approaches assume access to ground-truth captions. This assumption fails in realistic scenarios where only images are available and their textual annotations remain undisclosed, rendering prior methods ineffective when substituted with vision-language model (VLM) captions. In this work, we propose MoFit, a caption-free MIA framework that constructs synthetic conditioning inputs that are explicitly overfitted to the target model's generative manifold. Given a query image, MoFit proceeds in two stages: (i) model-fitted surrogate optimization, where a perturbation applied to the image is optimized to construct a surrogate in regions of the model's unconditional prior learned from member samples, and (ii) surrogate-driven embedding extraction, where a model-fitted embedding is derived from the surrogate and then used as a mismatched condition for the query image. This embedding amplifies conditional loss responses for member samples while leaving hold-outs relatively less affected, thereby enhancing separability in the absence of ground-truth captions. Our comprehensive experiments across multiple datasets and diffusion models demonstrate that MoFit consistently outperforms prior VLM-conditioned baselines and achieves performance competitive with caption-dependent methods.

Woo Jae Kim, Kyu Beom Han, Yoonki Cho et al.

As Neural Radiance Fields (NeRFs) have emerged as a powerful tool for 3D scene representation and novel view synthesis, protecting their intellectual property (IP) from unauthorized use is becoming increasingly crucial. In this work, we aim to protect the IP of NeRFs by injecting adversarial perturbations that disrupt their unauthorized applications. However, perturbing the 3D geometry of NeRFs can easily deform the underlying scene structure and thus substantially degrade the rendering quality, which has led existing attempts to avoid geometric perturbations or restrict them to explicit spaces like meshes. To overcome this limitation, we introduce a learnable sensitivity to quantify the spatially varying impact of geometric perturbations on rendering quality. Building upon this, we propose AegisRF, a novel framework that consists of a Perturbation Field, which injects adversarial perturbations into the pre-rendering outputs (color and volume density) of NeRF models to fool an unauthorized downstream target model, and a Sensitivity Field, which learns the sensitivity to adaptively constrain geometric perturbations, preserving rendering quality while disrupting unauthorized use. Our experimental evaluations demonstrate the generalized applicability of AegisRF across diverse downstream tasks and modalities, including multi-view image classification and voxel-based 3D localization, while maintaining high visual fidelity. Codes are available at https://github.com/wkim97/AegisRF.

Jaehan Kim, Minkyoo Song, Seungwon Shin et al.

Recent large language models (LLMs) have increasingly adopted the Mixture-of-Experts (MoE) architecture for efficiency. MoE-based LLMs heavily depend on a superficial safety mechanism in which harmful inputs are routed safety-critical experts. However, our analysis reveals that routing decisions for harmful inputs drift significantly after fine-tuning, exposing a critical vulnerability to harmful fine-tuning (HFT) attacks. Existing defenses, primarily designed for monolithic LLMs, are less effective for MoE LLMs as they fail to prevent drift in harmful input routing. To address this limitation, we propose SafeMoE, a safe fine-tuning method tailored to MoE LLMs. SafeMoE directly mitigates routing drift by penalizing the gap between the routing weights of a fine-tuned model and those of the initial safety-aligned model, thereby preserving the safety-aligned routing of harmful inputs to safety-critical experts. Experiments on open-source MoE LLMs ranging from 7B to 141B parameters demonstrate that SafeMoE effectively mitigates HFT attacks, reducing the harmfulness score of OLMoE from 62.0 to 5.0, for example, while maintaining task utility within 1% degradation and incurring only 2% overhead. It significantly outperforms state-of-the-art defense methods for safeguarding LLM fine-tuning and remains effective in recent large-scale MoE LLMs such as gpt-oss and Llama 4. Our implementation is available at https://anonymous.4open.science/r/SafeMoE.

Leo Marchyok, Zachary Coalson, Sungho Keum et al.

Modern language models exhibit rich internal structure, yet little is known about how privacy-sensitive behaviors, such as personally identifiable information (PII) leakage, are represented and modulated within their hidden states. We present UniLeak, a mechanistic-interpretability framework that identifies universal activation directions: latent directions in a model's residual stream whose linear addition at inference time consistently increases the likelihood of generating PII across prompts. These model-specific directions generalize across contexts and amplify PII generation probability, with minimal impact on generation quality. UniLeak recovers such directions without access to training data or groundtruth PII, relying only on self-generated text. Across multiple models and datasets, steering along these universal directions substantially increases PII leakage compared to existing prompt-based extraction methods. Our results offer a new perspective on PII leakage: the superposition of a latent signal in the model's representations, enabling both risk amplification and mitigation.

Joonsung Jeon, Woo Jae Kim, Suhyeon Ha et al.

The outstanding capability of diffusion models in generating high-quality images poses significant threats when misused by adversaries. In particular, we assume malicious adversaries exploiting diffusion models for inpainting tasks, such as replacing a specific region with a celebrity. While existing methods for protecting images from manipulation in diffusion-based generative models have primarily focused on image-to-image and text-to-image tasks, the challenge of preventing unauthorized inpainting has been rarely addressed, often resulting in suboptimal protection performance. To mitigate inpainting abuses, we propose ADVPAINT, a novel defensive framework that generates adversarial perturbations that effectively disrupt the adversary's inpainting tasks. ADVPAINT targets the self- and cross-attention blocks in a target diffusion inpainting model to distract semantic understanding and prompt interactions during image generation. ADVPAINT also employs a two-stage perturbation strategy, dividing the perturbation region based on an enlarged bounding box around the object, enhancing robustness across diverse masks of varying shapes and sizes. Our experimental results demonstrate that ADVPAINT's perturbations are highly effective in disrupting the adversary's inpainting tasks, outperforming existing methods; ADVPAINT attains over a 100-point increase in FID and substantial decreases in precision.

Suyoung Lee, Wonho Song, Suman Jana et al.

Trigger set-based watermarking schemes have gained emerging attention as they provide a means to prove ownership for deep neural network model owners. In this paper, we argue that state-of-the-art trigger set-based watermarking algorithms do not achieve their designed goal of proving ownership. We posit that this impaired capability stems from two common experimental flaws that the existing research practice has committed when evaluating the robustness of watermarking algorithms: (1) incomplete adversarial evaluation and (2) overlooked adaptive attacks. We conduct a comprehensive adversarial evaluation of 11 representative watermarking schemes against six of the existing attacks and demonstrate that each of these watermarking schemes lacks robustness against at least two non-adaptive attacks. We also propose novel adaptive attacks that harness the adversary's knowledge of the underlying watermarking algorithm of a target model. We demonstrate that the proposed attacks effectively break all of the 11 watermarking schemes, consequently allowing adversaries to obscure the ownership of any watermarked model. We encourage follow-up studies to consider our guidelines when evaluating the robustness of their watermarking schemes via conducting comprehensive adversarial evaluation that includes our adaptive attacks to demonstrate a meaningful upper bound of watermark robustness.

Byunggill Joe, Jihun Hamm, Sung Ju Hwang et al.

Although deep neural networks have shown promising performances on various tasks, they are susceptible to incorrect predictions induced by imperceptibly small perturbations in inputs. A large number of previous works proposed to detect adversarial attacks. Yet, most of them cannot effectively detect them against adaptive whitebox attacks where an adversary has the knowledge of the model and the defense method. In this paper, we propose a new probabilistic adversarial detector motivated by a recently introduced non-robust feature. We consider the non-robust features as a common property of adversarial examples, and we deduce it is possible to find a cluster in representation space corresponding to the property. This idea leads us to probability estimate distribution of adversarial representations in a separate cluster, and leverage the distribution for a likelihood based adversarial detector.

Dongkwan Kim, Eunsoo Kim, Sang Kil Cha et al.

Binary code similarity analysis (BCSA) is widely used for diverse security applications, including plagiarism detection, software license violation detection, and vulnerability discovery. Despite the surging research interest in BCSA, it is significantly challenging to perform new research in this field for several reasons. First, most existing approaches focus only on the end results, namely, increasing the success rate of BCSA, by adopting uninterpretable machine learning. Moreover, they utilize their own benchmark, sharing neither the source code nor the entire dataset. Finally, researchers often use different terminologies or even use the same technique without citing the previous literature properly, which makes it difficult to reproduce or extend previous work. To address these problems, we take a step back from the mainstream and contemplate fundamental research questions for BCSA. Why does a certain technique or a certain feature show better results than the others? Specifically, we conduct the first systematic study on the basic features used in BCSA by leveraging interpretable feature engineering on a large-scale benchmark. Our study reveals various useful insights on BCSA. For example, we show that a simple interpretable model with a few basic features can achieve a comparable result to that of recent deep learning-based approaches. Furthermore, we show that the way we compile binaries or the correctness of underlying binary analysis tools can significantly affect the performance of BCSA. Lastly, we make all our source code and benchmark public and suggest future directions in this field to help further research.

Suyoung Lee, HyungSeok Han, Sang Kil Cha et al.

JavaScript (JS) engine vulnerabilities pose significant security threats affecting billions of web browsers. While fuzzing is a prevalent technique for finding such vulnerabilities, there have been few studies that leverage the recent advances in neural network language models (NNLMs). In this paper, we present Montage, the first NNLM-guided fuzzer for finding JS engine vulnerabilities. The key aspect of our technique is to transform a JS abstract syntax tree (AST) into a sequence of AST subtrees that can directly train prevailing NNLMs. We demonstrate that Montage is capable of generating valid JS tests, and show that it outperforms previous studies in terms of finding vulnerabilities. Montage found 37 real-world bugs, including three CVEs, in the latest JS engines, demonstrating its efficacy in finding JS engine bugs.