Despite its practical impact, closed-box fuzzing on web applications remains understudied. This paper investigates two fundamental limitations of closed-box web fuzzing: (1) limited input space exploration due to the lack of a feedback mechanism, and (2) ineffective exploitation strategies caused by the shallow vulnerability identification. We propose Zelda, a novel closed-box web fuzzer designed to address these limitations. Specifically, we infer feedback signals from web responses in a closed-box testing environment, thereby deriving a feedback mechanism to guide the fuzzing process. We then coordinate two distinct input generation strategies for path exploration and exploitation, based on the exploration stage, which facilitates both in-page code coverage and vulnerability identification. Our evaluation across 15 real-world applications and nine benchmark sets demonstrates that Zelda's feedback mechanism and strategies are effective in practical web vulnerability discovery. Zelda uncovered 182 vulnerabilities, outperforming six state-of-the-art web fuzzers. In the wild, Zelda further discovered previously unreported vulnerabilities that received 29 CVE assignments.

Online question-and-answer (Q&A) platforms facilitate knowledge exchange through posted questions and answers. Recent advances in large language models (LLMs) have shown their strong capability in generating high-quality answers, leading to a recent surge in LLM-generated answers (LGAs) on Q&A platforms. In this paper, we conduct an in-depth analysis of how LGAs affect Naver Knowledge iN, the most popular Q&A platform in South Korea. To this end, we implement nine state-of-the-art LLM-generated text (LGT) detection methods and evaluate their performance on answers collected from Naver Knowledge iN. We then build an ensemble detector by stacking the three best-performing LGT detection methods, achieving an AUC of 0.9987 with a false positive rate below 1%. Using this LGA detector, we identify 75,558 LGAs among 1.46M answers. We find that LGAs tend to be longer, use more punctuation marks, and exhibit higher lexical diversity. However, LGAs do not show clear differences in user reactions, such as upvotes, downvotes, or selection rates by questioners. We also find that LGAs are primarily intended for knowledge sharing rather than personal experiences sharing. Finally, we observe a shift in the Q&A platform: questions increasingly move from simple fact-seeking to those involving complex contexts and seeking personal opinions or past experiences.

Mixture-of-Experts (MoE) has emerged as a prominent architecture for scaling large language models (LLMs). In particular, leveraging readily available fine-tuned LLMs as experts provides an efficient and flexible approach to developing MoE LLMs. However, integrating highly capable but untrustworthy LLMs into an MoE system poses a significant safety risk, potentially compromising the overall safety of the MoE LLM system. To date, no study has explored how adversaries compromise an MoE LLM service by introducing a poisoned expert LLM. In this paper, we introduce MOEVIL, a novel expert poisoning attack designed to compromise the safety of MoE LLMs. We address the dissipation of harmful effects from a target expert within MoE systems by conducting harmful preference learning. Next, we strategically manipulate this expert's latent vector to deceive the gating networks. This manipulation indirectly steers routing decisions toward the poisoned expert when generating responses to harmful queries. MOEVIL demonstrates strong attack performance across diverse MoE configurations based on both Llama and Qwen LLMs, even when poisoning only a single expert. MOEVIL increases the harmfulness score from 0.58 to 79.42 in a Llama-based MoE LLM, outperforming existing harmful poisoning attacks. Furthermore, our results demonstrate that even safety alignment, when combined with an efficient MoE training strategy, fails to fully mitigate these risks. Our findings demonstrate the significant threat posed by harmful experts in MoE systems, underscoring the need for robust safety measures in MoE-based LLM development. Our implementation is available at https://github.com/jaehanwork/MoEvil.

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.

Extensive academic effort has been put into the development of effective machine learning models that block advertising and tracking service (ATS) content. These ATS blockers leverage various features from websites, such as structural, content, flow, and JavaScript features, to develop accurate and robust models. However, establishing the robustness of these ATS blockers to evasion attacks is largely understudied, particularly in practical scenarios in which an adversary generates a single and cost-effective universal perturbation that renders ATS detection across websites ineffective at scale.In this paper, we show that recent ATS blockers using machine learning are not robust to a universal adversarial attack. Specifically, we propose an auditing framework (YOPO) that enables one to generate a single adversarial perturbation in a cost-effective manner. Our framework casts the generation of a universal perturbation into an optimization problem in a principled way; it enables an adversary to minimize the cost of manipulating various features in HTML content and to thwart ATS classification while constraining the perturbation size for each feature. We demonstrate that YOPO is capable of generating a universal perturbation that enables bypassing four seminal ATS blockers: AdGraph, WebGraph, AdFlush, and PageGraph, attaining success rates of up to 92.27%, 71.50%, 61.91%, and 85.81%, respectively. We also propose a practical and effective countermeasure against YOPO that only requires preprocessing training instances without large performance drops in ATS blocking.

Advertising and tracking service (ATS) blocking has been safeguarding the privacy of millions of Internet users from privacy-invasive tracking behaviors. Previous research has proposed using a graph representation that models the structural relationships in loading web resources and then conducting ATS node classification based on this graph representation. However, these context-based ATS classification methods suffer from (1) inconsistent classification due to the varying context in which ATS resources are loaded and (2) a lack of explainability of the classification results, making it difficult to identify the code-level causes for ATS classification.

As Machine Learning as a Service (MLaaS) platforms become prevalent, deep neural network (DNN) watermarking techniques are gaining increasing attention, which enables one to verify the ownership of a target DNN model in a black-box scenario. Unfortunately, previous watermarking methods are vulnerable to functionality stealing attacks, thus allowing an adversary to falsely claim the ownership of a DNN model stolen from its original owner. In this work, we propose a novel margin-based DNN watermarking approach that is robust to the functionality stealing attacks based on model extraction and distillation. Specifically, during training, our method maximizes the margins of watermarked samples by using projected gradient ascent on them so that their predicted labels cannot change without compromising the accuracy of the model that the attacker tries to steal. We validate our method on multiple benchmarks and show that our watermarking method successfully defends against model extraction attacks, outperforming recent baselines.

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.

A Sybil attack is a critical threat that undermines the trust and integrity of web services by creating and exploiting a large number of fake (i.e., Sybil) accounts. To mitigate this threat, previous studies have proposed leveraging collective classification to detect Sybil accounts. Recently, researchers have demonstrated that state-of-the-art adversarial attacks are able to bypass existing collective classification methods, posing a new security threat. To this end, we propose RICC, the first robust collective classification framework, designed to identify adversarial Sybil accounts created by adversarial attacks. RICC leverages the novel observation that these adversarial attacks are highly tailored to a target collective classification model to optimize the attack budget. Owing to this adversarial strategy, the classification results for adversarial Sybil accounts often significantly change when deploying a new training set different from the original training set used for assigning prior reputation scores to user accounts. Leveraging this observation, RICC achieves robustness in collective classification by stabilizing classification results across different training sets randomly sampled in each round. RICC achieves false negative rates of 0.01, 0.11, 0.00, and 0.01 in detecting adversarial Sybil accounts for the Enron, Facebook, Twitter_S, and Twitter_L datasets, respectively. It also attains respective AUCs of 0.99, 1.00, 0.89, and 0.74 for these datasets, achieving high performance on the original task of detecting Sybil accounts. RICC significantly outperforms all existing Sybil detection methods, demonstrating superior robustness and efficacy in the collective classification of Sybil accounts.

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.

Hardware security modules (HSMs) have been utilized as a trustworthy foundation for cloud services. Unfortunately, existing systems using HSMs fail to meet multi-tenant scalability arising from the emerging trends such as microservices, which utilize frequent cryptographic operations. As an alternative, cloud vendors provide HSMs as a service. However, such cloud-managed HSM usage models raise security concerns due to their untrusted and shared operating environment. We propose ScaleTrust, a scalable and secure system for key management. ScaleTrust allows us to scale the number of virtual HSM partitions, each of which is isolated with respect to each other and is robust against cloud insider attacks, while preserving physical isolation of the root of trust. To enable this, ScaleTrust uses Intel SGX and multiple HSM features, such as restricting key usage by controlling key attributes of in-HSM keys and establishing a secure channel using only HSM commands. Finally, we apply ScaleTrust to four real-world systems: Keyless SSL for TLS private key offloading, JSON Web Token authentication for microservices, key provisioning, and encryption in database systems. Our evaluation shows that ScaleTrust achieves multi-tenancy in a scalable way by providing multiple virtual HSMs with legacy HSM devices that are designed to support a single tenant. ScaleTrust provides security against insider threats while incurring 11.9% and 39.0% of end-to-end throughput and latency overhead for Keyless SSL compared to stand-alone HSMs.

Binary code similarity analysis (BCSA) is widely used for diverse security applications such as 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 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.

In South Korea, voice phishing has been proliferating with the advent of voice phishing apps: the number of annual victims had risen to 34,527 in 2020, representing financial losses of approximately 598 million USD. However, the voice phishing functionalities that these abusive apps implement are largely understudied. To this end, we analyze 1,017 voice phishing apps and reveal new phishing functionalities: outgoing call redirection, call screen overlay, and fake call voice. We find that call redirection that changes the intended recipients of victims' outgoing calls plays a critical role in facilitating voice phishing; our user study shows that 87% of the participants did not notice that their intended recipients were changed when call redirection occurred. We further investigate implementations of these fatal functionalities to distinguish their malicious behaviors from their corresponding behaviors in benign apps. We then propose HearMeOut, an Android system-level service that detects phishing behaviors that phishing apps conduct in runtime and blocks the detected behaviors. HearMeOut achieves high accuracy with no false positives or negatives in classifying phishing behaviors while exhibiting an unnoticeable latency of 0.36 ms on average. Our user study demonstrates that HearMeOut is able to prevent 100% of participants from being phished by providing active warnings. Our work facilitates a better understanding of recent voice phishing and proposes practical mitigation with recommendations for Android system changes.

The wide adoption of deep neural networks (DNNs) in mission-critical applications has spurred the need for interpretable models that provide explanations of the model's decisions. Unfortunately, previous studies have demonstrated that model explanations facilitate information leakage, rendering DNN models vulnerable to model inversion attacks. These attacks enable the adversary to reconstruct original images based on model explanations, thus leaking privacy-sensitive features. To this end, we present Generative Noise Injector for Model Explanations (GNIME), a novel defense framework that perturbs model explanations to minimize the risk of model inversion attacks while preserving the interpretabilities of the generated explanations. Specifically, we formulate the defense training as a two-player minimax game between the inversion attack network on the one hand, which aims to invert model explanations, and the noise generator network on the other, which aims to inject perturbations to tamper with model inversion attacks. We demonstrate that GNIME significantly decreases the information leakage in model explanations, decreasing transferable classification accuracy in facial recognition models by up to 84.8% while preserving the original functionality of model explanations.

Black-box web scanners have been a prevalent means of performing penetration testing to find reflected cross-site scripting (XSS) vulnerabilities. Unfortunately, off-the-shelf black-box web scanners suffer from unscalable testing as well as false negatives that stem from a testing strategy that employs fixed attack payloads, thus disregarding the exploitation of contexts to trigger vulnerabilities. To this end, we propose a novel method of adapting attack payloads to a target reflected XSS vulnerability using reinforcement learning (RL). We present Link, a general RL framework whose states, actions, and a reward function are designed to find reflected XSS vulnerabilities in a black-box and fully automatic manner. Link finds 45, 213, and 60 vulnerabilities with no false positives in Firing-Range, OWASP, and WAVSEP benchmarks, respectively, outperforming state-of-the-art web scanners in terms of finding vulnerabilities and ending testing campaigns earlier. Link also finds 43 vulnerabilities in 12 real-world applications, demonstrating the promising efficacy of using RL in finding reflected XSS vulnerabilities.

A code property graph (CPG) is a joint representation of syntax, control flows, and data flows of a target application. Recent studies have demonstrated the promising efficacy of leveraging CPGs for the identification of vulnerabilities. It recasts the problem of implementing a specific static analysis for a target vulnerability as a graph query composition problem. It requires devising coarse-grained graph queries that model vulnerable code patterns. Unfortunately, such coarse-grained queries often leave vulnerabilities due to faulty input sanitization undetected. In this paper, we propose, a scalable system designed to identify various web vulnerabilities, including bugs that stem from incorrect sanitization. We designed to find a subgraph in a target CPG that matches a given CPG query having a known vulnerability, which is known as the subgraph isomorphism problem. To address the scalability challenge that stems from the NP-complete nature of this problem, leverages optimization techniques designed to boost the efficiency of matching vulnerable subgraphs. found confirmed vulnerabilities including CVEs among 2,464 potential vulnerabilities in real-world CPGs having a combined total of 1 billion nodes and 1.2 billion edges.

Phishing attacks are causing substantial damage albeit extensive effort in academia and industry. Recently, a large volume of phishing attacks transit toward adopting HTTPS, leveraging TLS certificates issued from Certificate Authorities (CAs), to make the attacks more effective. In this paper, we present a comprehensive study on the security practices of CAs in the HTTPS phishing ecosystem. We focus on the CAs, critical actors under-studied in previous literature, to better understand the importance of the security practices of CAs and thwart the proliferating HTTPS phishing. In particular, we first present the current landscape and effectiveness of HTTPS phishing attacks comparing to traditional HTTP ones. Then, we conduct an empirical experiment on the CAs' security practices in terms of the issuance and revocation of the certificates. Our findings highlight serious conflicts between the expected security practices of CAs and reality, raising significant security concerns. We further validate our findings using a longitudinal dataset of abusive certificates used for real phishing attacks in the wild. We confirm that the security concerns of CAs prevail in the wild and these concerns can be one of the main contributors to the recent surge of HTTPS phishing attacks.

The era of Smart Homes and the Internet of Things (IoT) calsl for integrating diverse "smart" devices, including sensors, actuators, and home appliances. However, enabling interoperation across heterogeneous IoT devices is a challenging task because vendors use their own control and communication protocols. Prior approaches have attempted to solve this problem by asking for vendor support, or even fundamentally re-designing the architecture of IoT devices. These approaches face limitations as they require disruptive changes.This paper explores a new approach to improving IoT interoperability without requiring architectural changes or vendor participation. Focusing on smart-home environments, we propose Lumos that improves interoperability by leveraging Android apps that control IoT devices. Lumos uses this information learned from IoT apps to enable "best-effort" interoperation across heterogeneous devices. Our evaluation with 15 commercial IoT devices from three major IoT platforms and in-depth user studies conducted with 24 participants demonstrate the promising efficacy of Lumos for implementing diverse interoperation scenarios.

The early discovery of security bugs in JavaScript (JS) engines is crucial for protecting Internet users from adversaries abusing zero-day vulnerabilities. Browser vendors, bug bounty hunters, and security researchers have been eager to find such security bugs by leveraging state-of-the-art fuzzers as well as their domain expertise. They report a bug when observing a crash after executing their JS test since a crash is an early indicator of a potential bug. However, it is difficult to identify whether such a crash indeed invokes security bugs in JS engines. Thus, unskilled bug reporters are unable to assess the security severity of their new bugs with JS engine crashes. Today, this classification of a reported security bug is completely manual, depending on the verdicts from JS engine vendors. We investigated the feasibility of applying various machine learning classifiers to determine whether an observed crash triggers a security bug. We designed and implemented CRScope, which classifies security and non-security bugs from given crash-dump files. Our experimental results on 766 crash instances demonstrate that CRScope achieved 0.85, 0.89, and 0.93 Area Under Curve (AUC) for Chakra, V8, and SpiderMonkey crashes, respectively. CRScope also achieved 0.84, 0.89, and 0.95 precision for Chakra, V8, and SpiderMonkey crashes, respectively. This outperforms the previous study and existing tools including Exploitable and AddressSanitizer. CRScope is capable of learning domain-specific expertise from the past verdicts on reported bugs and automatically classifying JS engine security bugs, which helps improve the scalable classification of security bugs.

Anonymous network services on the World Wide Web have emerged as a new web architecture, called the Dark Web. The Dark Web has been notorious for harboring cybercriminals abusing anonymity. At the same time, the Dark Web has been a last resort for people who seek freedom of the press as well as avoid censorship. This anonymous nature allows website operators to conceal their identity and thereby leads users to have difficulties in determining the authenticity of websites. Phishers abuse this perplexing authenticity to lure victims; however, only a little is known about the prevalence of phishing attacks on the Dark Web.

Progressive Web App (PWA) is a new generation of Web application designed to provide native app-like browsing experiences even when a browser is offline. PWAs make full use of new HTML5 features which include push notification, cache, and service worker to provide short-latency and rich Web browsing experiences. We conduct the first systematic study of the security and privacy aspects unique to PWAs. We identify security flaws in main browsers as well as design flaws in popular third-party push services, that exacerbate the phishing risk. We introduce a new side-channel attack that infers the victim's history of visited PWAs. The proposed attack exploits the offline browsing feature of PWAs using a cache. We demonstrate a cryptocurrency mining attack which abuses service workers. Defenses and recommendations to mitigate the identified security and privacy risks are suggested with in-depth understanding.

Code injection attacks continue to plague applications that incorporate user input into executable programs. For example, SQL injection vulnerabilities rank fourth among all bugs reported in CVE, yet all previously proposed methods for detecting SQL injection attacks suffer from false positives and false negatives.

The OpenFlow (OF) switching specification represents an innovative and open standard for enabling the dynamic programming of flow control policies in production networks. Unfortunately, thus far researchers have paid little attention to the development of methods for verifying that dynamic flow policies inserted within an OpenFlow network do not violate the network's underlying security policy. We introduce Flover, a model checking system which verifies that the aggregate of flow policies instantiated within an OpenFlow network does not violate the network's security policy. We have implemented Flover using the Yices SMT solver, which we then integrated into NOX, a popular OpenFlow network controller. Flover provides NOX a formal validation of the OpenFlow network's security posture.

Web applications written in languages such as PHP and JSP are notoriously vulnerable to accidentally omitted authorization checks and other security bugs. Existing techniques that find missing security checks in library and system code assume that (1) security checks can be recognized syntactically and (2) the same pattern of checks applies universally to all programs. These assumptions do not hold for Web applications. Each Web application uses different variables and logic to check the user's permissions. Even within the application, security logic varies based on the user's role, e.g., regular users versus administrators. This paper describes ROLECAST, the first system capable of statically identifying security logic that mediates security-sensitive events (such as database writes) in Web applications, rather than taking a specification of this logic as input. We observe a consistent software engineering pattern-the code that implements distinct user role functionality and its security logic resides in distinct methods and files-and develop a novel algorithm for discovering this pattern in Web applications. Our algorithm partitions the set of file contexts (a coarsening of calling contexts) on which security-sensitive events are control dependent into roles. Roles are based on common functionality and security logic. ROLECAST identifies security-critical variables and applies rolespecific variable consistency analysis to find missing security checks. ROLECAST discovered 13 previously unreported, remotely exploitable vulnerabilities in 11 substantial PHP and JSP applications, with only 3 false positives.