Jaechul Roh

Jaechul Roh, Minhao Cheng, Yajun Fang

Pre-trained language models allowed us to process downstream tasks with the help of fine-tuning, which aids the model to achieve fairly high accuracy in various Natural Language Processing (NLP) tasks. Such easily-downloaded language models from various websites empowered the public users as well as some major institutions to give a momentum to their real-life application. However, it was recently proven that models become extremely vulnerable when they are backdoor attacked with trigger-inserted poisoned datasets by malicious users. The attackers then redistribute the victim models to the public to attract other users to use them, where the models tend to misclassify when certain triggers are detected within the training sample. In this paper, we will introduce a novel improved textual backdoor defense method, named MSDT, that outperforms the current existing defensive algorithms in specific datasets. The experimental results illustrate that our method can be effective and constructive in terms of defending against backdoor attack in text domain. Code is available at https://github.com/jcroh0508/MSDT.

Jaechul Roh, Eugene Bagdasarian, Hamed Haddadi et al.

LLM-powered agents are beginning to automate user's tasks across the open web, often with access to user resources such as emails and calendars. Unlike standard LLMs answering questions in a controlled ChatBot setting, web agents act "in the wild", interacting with third parties and leaving behind an action trace. Therefore, we ask the question: how do web agents handle user resources when accomplishing tasks on their behalf across live websites? In this paper, we formalize Natural Agentic Oversharing -- the unintentional disclosure of task-irrelevant user information through an agent trace of actions on the web. We introduce SPILLage, a framework that characterizes oversharing along two dimensions: channel (content vs. behavior) and directness (explicit vs. implicit). This taxonomy reveals a critical blind spot: while prior work focuses on text leakage, web agents also overshare behaviorally through clicks, scrolls, and navigation patterns that can be monitored. We benchmark 180 tasks on live e-commerce sites with ground-truth annotations separating task-relevant from task-irrelevant attributes. Across 1,080 runs spanning two agentic frameworks and three backbone LLMs, we demonstrate that oversharing is pervasive with behavioral oversharing dominates content oversharing by 5x. This effect persists -- and can even worsen -- under prompt-level mitigation. However, removing task-irrelevant information before execution improves task success by up to 17.9%, demonstrating that reducing oversharing improves task success. Our findings underscore that protecting privacy in web agents is a fundamental challenge, requiring a broader view of "output" that accounts for what agents do on the web, not just what they type. Our datasets and code are available at https://github.com/jrohsc/SPILLage.

Jaechul Roh, Jean-Philippe Monteuuis, Jonathan Petit et al.

Prior attacks on Audio Large Language Models (Audio LLMs) demonstrated that carefully crafted waveform-domain perturbations can force targeted adversarial outputs. As a defense mechanism against these attacks, real-world codec compression preprocessing has been studied to both detect and remove the perturbations. Yet no existing attack has demonstrated robustness against these compressions. We introduce CodecAttack, which optimizes a perturbation in a neural audio codec's continuous latent space rather than directly perturbing the audio waveform. We show that the codec's compression channel, which discards waveform perturbations, transmits perturbations crafted in its own latent space. To further harden the attack across real-world compression channels, we apply multi-bitrate straight-through Expectation-over-Transformation (EoT), all without modifying the target model. Across three realistic Audio LLM deployment scenarios and three target models, CodecAttack achieves an average 85.5% target-substring attack success rate (ASR) on Opus at moderate bitrates, while the waveform baseline trained with identical EoT hardening does not exceed 26% at any bitrate. The attack transfers to held-out codecs, reaching up to 100% ASR on MP3 and 84% on AAC-LC without retraining. A per-band energy analysis shows that the latent perturbation concentrates below 4kHz, exactly where codecs allocate the most bits, while the waveform baseline spreads into higher frequencies that codecs discard. These results demonstrate that lossy compression is not a reliable defense against adversarial audio and that codec-aware attacks pose a practical threat to deployed Audio LLM systems.

Jaechul Roh, Yajun Fang

Over the past few years, the field of adversarial attack received numerous attention from various researchers with the help of successful attack success rate against well-known deep neural networks that were acknowledged to achieve high classification ability in various tasks. However, majority of the experiments were completed under a single model, which we believe it may not be an ideal case in a real-life situation. In this paper, we introduce a novel federated adversarial training method for smart home face recognition, named FLATS, where we observed some interesting findings that may not be easily noticed in a traditional adversarial attack to federated learning experiments. By applying different variations to the hyperparameters, we have spotted that our method can make the global model to be robust given a starving federated environment. Our code can be found on https://github.com/jcroh0508/FLATS.

Jaechul Roh, Amir Houmansadr

Prior work shows that fine-tuning aligned models on benign data degrades safety in text and vision modalities, and that proximity to harmful content in representation space predicts which samples cause the most damage. However, existing analyses operate within a single, undifferentiated embedding space -- leaving open whether distinct input properties drive the vulnerability differently. Audio introduces a structurally richer problem: a benign sample can neighbor harmful content not only through what is said but through how it sounds, even when its words are entirely innocuous. We present the first systematic study of benign fine-tuning safety in Audio LLMs, evaluating three state-of-the-art models with a proximity-based filtering framework that selects benign audio by embedding-space distance to harmful content. By decomposing proximity into semantic, acoustic, and mixed axes using external reference encoders alongside each model's own internal encoder, we show that benign fine-tuning elevates Jailbreak Success Rate (JSR) from single digits to as high as 87.12%. Crucially, the dominant vulnerability axis and the relative risk of audio versus text fine-tuning are both architecture-conditioned -- determined by how each model's encoder and projector transform audio into the LLM's input space. We propose two defenses: filtering training data to maximize distance from harmful embeddings, and a textual system prompt at inference, both reducing JSR to near-zero without architectural modification. Our mechanistic analysis on two architectures reveals that fine-tuning selectively suppresses the late-layer refusal circuit while the frozen encoder preserves representations, and that even the suppression pattern is architecture-conditioned, mirroring the behavioral asymmetries across modalities. Safety degradation from benign fine-tuning is a qualitatively distinct risk in Audio LLMs.

Moo Hyun Son, Jintaek Oh, Sun Bin Mun et al.

While text-to-image (T2I) models can synthesize high-quality images, their performance degrades significantly when prompted with novel or out-of-distribution (OOD) entities due to inherent knowledge cutoffs. We introduce World-To-Image, a novel framework that bridges this gap by empowering T2I generation with agent-driven world knowledge. We design an agent that dynamically searches the web to retrieve images for concepts unknown to the base model. This information is then used to perform multimodal prompt optimization, steering powerful generative backbones toward an accurate synthesis. Critically, our evaluation goes beyond traditional metrics, utilizing modern assessments like LLMGrader and ImageReward to measure true semantic fidelity. Our experiments show that World-To-Image substantially outperforms state-of-the-art methods in both semantic alignment and visual aesthetics, achieving +8.1% improvement in accuracy-to-prompt on our curated NICE benchmark. Our framework achieves these results with high efficiency in less than three iterations, paving the way for T2I systems that can better reflect the ever-changing real world. Our demo code is available here\footnote{https://github.com/mhson-kyle/World-To-Image}.

Abhinav Kumar, Jaechul Roh, Ali Naseh et al.

We increase overhead for applications that rely on reasoning LLMs-we force models to spend an amplified number of reasoning tokens, i.e., "overthink", to respond to the user query while providing contextually correct answers. The adversary performs an OVERTHINK attack by injecting decoy reasoning problems into the public content that is used by the reasoning LLM (e.g., for RAG applications) during inference time. Due to the nature of our decoy problems (e.g., a Markov Decision Process), modified texts do not violate safety guardrails. We evaluated our attack across closed-(OpenAI o1, o1-mini, o3-mini) and open-(DeepSeek R1) weights reasoning models on the FreshQA and SQuAD datasets. Our results show up to 18x slowdown on FreshQA dataset and 46x slowdown on SQuAD dataset. The attack also shows high transferability across models. To protect applications, we discuss and implement defenses leveraging LLM-based and system design approaches. Finally, we discuss societal, financial, and energy impacts of OVERTHINK attack which could amplify the costs for third-party applications operating reasoning models.

Jaechul Roh, Virat Shejwalkar, Amir Houmansadr

Large Audio Language Models (LALMs) have significantly advanced audio understanding but introduce critical security risks, particularly through audio jailbreaks. While prior work has focused on English-centric attacks, we expose a far more severe vulnerability: adversarial multilingual and multi-accent audio jailbreaks, where linguistic and acoustic variations dramatically amplify attack success. In this paper, we introduce Multi-AudioJail, the first systematic framework to exploit these vulnerabilities through (1) a novel dataset of adversarially perturbed multilingual/multi-accent audio jailbreaking prompts, and (2) a hierarchical evaluation pipeline revealing that how acoustic perturbations (e.g., reverberation, echo, and whisper effects) interacts with cross-lingual phonetics to cause jailbreak success rates (JSRs) to surge by up to +57.25 percentage points (e.g., reverberated Kenyan-accented attack on MERaLiON). Crucially, our work further reveals that multimodal LLMs are inherently more vulnerable than unimodal systems: attackers need only exploit the weakest link (e.g., non-English audio inputs) to compromise the entire model, which we empirically show by multilingual audio-only attacks achieving 3.1x higher success rates than text-only attacks. We plan to release our dataset to spur research into cross-modal defenses, urging the community to address this expanding attack surface in multimodality as LALMs evolve.

Ali Naseh, Jaechul Roh, Amir Houmansadr

Multimodal machine learning, especially text-to-image models like Stable Diffusion and DALL-E 3, has gained significance for transforming text into detailed images. Despite their growing use and remarkable generative capabilities, there is a pressing need for a detailed examination of these models' behavior, particularly with respect to memorization. Historically, memorization in machine learning has been context-dependent, with diverse definitions emerging from classification tasks to complex models like Large Language Models (LLMs) and Diffusion models. Yet, a definitive concept of memorization that aligns with the intricacies of text-to-image synthesis remains elusive. This understanding is vital as memorization poses privacy risks yet is essential for meeting user expectations, especially when generating representations of underrepresented entities. In this paper, we introduce a specialized definition of memorization tailored to text-to-image models, categorizing it into three distinct types according to user expectations. We closely examine the subtle distinctions between intended and unintended memorization, emphasizing the importance of balancing user privacy with the generative quality of the model outputs. Using the Stable Diffusion model, we offer examples to validate our memorization definitions and clarify their application.

Ali Naseh, Jaechul Roh, Amir Houmansadr

Diffusion-based models, such as the Stable Diffusion model, have revolutionized text-to-image synthesis with their ability to produce high-quality, high-resolution images. These advancements have prompted significant progress in image generation and editing tasks. However, these models also raise concerns due to their tendency to memorize and potentially replicate exact training samples, posing privacy risks and enabling adversarial attacks. Duplication in training datasets is recognized as a major factor contributing to memorization, and various forms of memorization have been studied so far. This paper focuses on two distinct and underexplored types of duplication that lead to replication during inference in diffusion-based models, particularly in the Stable Diffusion model. We delve into these lesser-studied duplication phenomena and their implications through two case studies, aiming to contribute to the safer and more responsible use of generative models in various applications.

Ali Naseh, Harsh Chaudhari, Jaechul Roh et al.

DeepSeek recently released R1, a high-performing large language model (LLM) optimized for reasoning tasks. Despite its efficient training pipeline, R1 achieves competitive performance, even surpassing leading reasoning models like OpenAI's o1 on several benchmarks. However, emerging reports suggest that R1 refuses to answer certain prompts related to politically sensitive topics in China. While existing LLMs often implement safeguards to avoid generating harmful or offensive outputs, R1 represents a notable shift - exhibiting censorship-like behavior on politically charged queries. In this paper, we investigate this phenomenon by first introducing a large-scale set of heavily curated prompts that get censored by R1, covering a range of politically sensitive topics, but are not censored by other models. We then conduct a comprehensive analysis of R1's censorship patterns, examining their consistency, triggers, and variations across topics, prompt phrasing, and context. Beyond English-language queries, we explore censorship behavior in other languages. We also investigate the transferability of censorship to models distilled from the R1 language model. Finally, we propose techniques for bypassing or removing this censorship. Our findings reveal possible additional censorship integration likely shaped by design choices during training or alignment, raising concerns about transparency, bias, and governance in language model deployment.

Jaechul Roh, Zachary Novack, Yuefeng Peng et al.

Generative AI systems for music and video commonly use text-based filters to prevent the regurgitation of copyrighted material. We expose a fundamental flaw in this approach by introducing Adversarial PhoneTic Prompting (APT), a novel attack that bypasses these safeguards by exploiting phonetic memorization. The APT attack replaces iconic lyrics with homophonic but semantically unrelated alternatives (e.g., "mom's spaghetti" becomes "Bob's confetti"), preserving acoustic structure while altering meaning; we identify high-fidelity phonetic matches using CMU pronouncing dictionary. We demonstrate that leading Lyrics-to-Song (L2S) models like SUNO and YuE regenerate songs with striking melodic and rhythmic similarity to their copyrighted originals when prompted with these altered lyrics. More surprisingly, this vulnerability extends across modalities. When prompted with phonetically modified lyrics from a song, a Text-to-Video (T2V) model like Veo 3 reconstructs visual scenes from the original music video-including specific settings and character archetypes-despite the absence of any visual cues in the prompt. Our findings reveal that models memorize deep, structural patterns tied to acoustics, not just verbatim text. This phonetic-to-visual leakage represents a critical vulnerability in transcript-conditioned generative models, rendering simple copyright filters ineffective and raising urgent concerns about the secure deployment of multimodal AI systems. Demo examples are available at our project page (https://jrohsc.github.io/music_attack/).

Jaechul Roh, Varun Gandhi, Shivani Anilkumar et al.

Large Language Models (LLMs) have achieved remarkable success in tasks requiring complex reasoning, such as code generation, mathematical problem solving, and algorithmic synthesis -- especially when aided by reasoning tokens and Chain-of-Thought prompting. Yet, a core question remains: do these models truly reason, or do they merely exploit shallow statistical patterns? In this paper, we introduce Chain-of-Code Collapse, where we systematically investigate the robustness of reasoning LLMs by introducing a suite of semantically faithful yet adversarially structured prompt perturbations. Our evaluation -- spanning 700 perturbed code generations derived from LeetCode-style problems -- applies transformations such as storytelling reframing, irrelevant constraint injection, example reordering, and numeric perturbation. We observe that while certain modifications severely degrade performance (with accuracy drops up to -42.1%), others surprisingly improve model accuracy by up to 35.3%, suggesting sensitivity not only to semantics but also to surface-level prompt dynamics. These findings expose the fragility and unpredictability of current reasoning systems, underscoring the need for more principles approaches to reasoning alignments and prompting robustness. We release our perturbation datasets and evaluation framework to promote further research in trustworthy and resilient LLM reasoning.

Abhinav Kumar, Jaechul Roh, Ali Naseh et al.

AI web agents use Internet resources at far greater speed, scale, and complexity -- changing how users and services interact. Deployed maliciously or erroneously, these agents could overload content providers. At the same time, web agents can bypass CAPTCHAs and other defenses by mimicking user behavior or flood authentication systems with fake accounts. Yet providers must protect their services and content from denial-of-service attacks and scraping by web agents. In this paper, we design a framework that imposes tunable costs on agents before providing access to resources; we call this Web Agent Throttling. We start by formalizing Throttling Gates as challenges issued to an agent that are asymmetric, scalable, robust, and compatible with any agent. Focusing on a common component -- the language model -- we require the agent to solve reasoning puzzles, thereby incurring excessive token-generation costs. However, we find that using existing puzzles, e.g., coding or math, as throttling gates fails to satisfy our properties. To address this, we introduce rebus-based Reasoning Gates, synthetic text puzzles that require multi-hop reasoning over world knowledge (thereby throttling an agent's model). We design a scalable generation and verification protocol for such reasoning gates. Our framework achieves computational asymmetry, i.e., the response-generation cost is 9.2x higher than the generation cost for SOTA models. We further deploy reasoning gates on a custom website and Model Context Protocol (MCP) servers and evaluate with real-world web agents. Finally, we discuss the limitations and environmental impact of real-world deployment of our framework.

Jaechul Roh, Andrew Yuan, Jinsong Mao

Text-to-Image (T2I) diffusion models have rapidly advanced, enabling the generation of high-quality images that align closely with textual descriptions. However, this progress has also raised concerns about their misuse for propaganda and other malicious activities. Recent studies reveal that attackers can embed biases into these models through simple fine-tuning, causing them to generate targeted imagery when triggered by specific phrases. This underscores the potential for T2I models to act as tools for disseminating propaganda, producing images aligned with an attacker's objective for end-users. Building on this concept, we introduce FameBias, a T2I biasing attack that manipulates the embeddings of input prompts to generate images featuring specific public figures. Unlike prior methods, Famebias operates solely on the input embedding vectors without requiring additional model training. We evaluate FameBias comprehensively using Stable Diffusion V2, generating a large corpus of images based on various trigger nouns and target public figures. Our experiments demonstrate that FameBias achieves a high attack success rate while preserving the semantic context of the original prompts across multiple trigger-target pairs.

Ali Naseh, Jaechul Roh, Eugene Bagdasarian et al.

Recent advances in large text-conditional diffusion models have revolutionized image generation by enabling users to create realistic, high-quality images from textual prompts, significantly enhancing artistic creation and visual communication. However, these advancements also introduce an underexplored attack opportunity: the possibility of inducing biases by an adversary into the generated images for malicious intentions, e.g., to influence public opinion and spread propaganda. In this paper, we study an attack vector that allows an adversary to inject arbitrary bias into a target model. The attack leverages low-cost backdooring techniques using a targeted set of natural textual triggers embedded within a small number of malicious data samples produced with public generative models. An adversary could pick common sequences of words that can then be inadvertently activated by benign users during inference. We investigate the feasibility and challenges of such attacks, demonstrating how modern generative models have made this adversarial process both easier and more adaptable. On the other hand, we explore various aspects of the detectability of such attacks and demonstrate that the model's utility remains intact in the absence of the triggers. Our extensive experiments using over 200,000 generated images and against hundreds of fine-tuned models demonstrate the feasibility of the presented backdoor attack. We illustrate how these biases maintain strong text-image alignment, highlighting the challenges in detecting biased images without knowing that bias in advance. Our cost analysis confirms the low financial barrier (\$10-\$15) to executing such attacks, underscoring the need for robust defensive strategies against such vulnerabilities in diffusion models.