arXiv subjects

Amit Dvir

Publications and source records attributed to Amit Dvir.

At least 19 recordsLinked to original sources

arXiv2026

Signing the Transaction but Not the Decision: Whisper Attacks and a Binding Defense for AP2

Software agents are beginning to shop and pay on a person's behalf. Agent payment protocols such as AP2 produce cryptographically valid signatures for completed purchases, yet do not constrain the decisions that lead to them. Consequently, ordinary product-description text can steer a shopping agent into forming a cart that passes every protocol check but no longer matches the user's request. In this paper, we show that this vulnerability enables three related attacks. In the first attack, the agent is steered into fetching another user's payment credentials. In the second, it assembles a cryptographically valid cart whose contents do not match what the user was shown. In the third, a single factual claim about stock or product lineage moves the agent from the cheaper displayed item to a more expensive one, while the resulting cart remains fully consistent with the listing. In experiments using the Gemini Flash-Lite models that AP2's sample agents specify by default, the three attacks succeeded at rates of 90%, 56%, and 73.3%, respectively. The same vulnerability appears across seventeen Google models, three unrelated agent frameworks, two cross-vendor anchors, and Google's own consumer assistant. To address this attack vector, we introduce A-VIP (AP2 Verified-Intent Protection), a protocol-layer defense that treats the signed intent as a capability grant rather than judging the merchant's description. The defense binds every credential lookup to the session that requested it and every cart line to the listing seen, while flagging unauthorized spending. The first two attacks leave structural traces that these bindings block with zero false positives. The third attack leaves no trace, so A-VIP surfaces unauthorized spending for user confirmation. Finally, we release the A-VIP code, machine-checked invariants, and AP2-WhisperBench, a suite of 1,544 evaluation scenarios.

cs.CR↗

arXiv2026

HTTP REST API Structure Learning

Application Programming Interfaces (APIs) are essential in software development, enabling web services, mobile apps, and microservices. However, their widespread use introduces significant security risks, highlighting the importance of API security. This paper presents HTTP REST API Learning (HRAL), a novel unsupervised anomaly detection approach that models the structure and behavior of API endpoints directly from network traffic, without relying on predefined rules or documentation. HRAL enables robust detection of malicious activity by understanding how APIs behave and flagging deviations as potential threats. We evaluate HRAL across varying levels of OpenAPI documentation detail and compare it with existing techniques. HRAL achieves strong performance, with an average recall of 82.07% and an F1-score of 87.24%, significantly outperforming alternatives when API documentation is limited. Moreover, our results approach the effectiveness of full API document definitions. When combined with signature-based rules such as the OWASP ModSecurity CRS, our system achieves 100% detection. These results highlight HRAL's effectiveness in real-world, partially documented API environments and its potential as a foundational layer for modern API security solutions.

cs.SE↗

arXiv2025

Security Analysis of Agentic AI Communication Protocols: A Comparative Evaluation

Multi-agent systems (MAS) powered by artificial intelligence (AI) are increasingly foundational to complex, distributed workflows. Yet, the security of their underlying communication protocols remains critically under-examined. This paper presents the first empirical, comparative security analysis of the official CORAL implementation and a high-fidelity, SDK-based ACP implementation, benchmarked against a literature-based evaluation of A2A. Using a 14 point vulnerability taxonomy, we systematically assess their defenses across authentication, authorization, integrity, confidentiality, and availability. Our results reveal a pronounced security dichotomy: CORAL exhibits a robust architectural design, particularly in its transport-layer message validation and session isolation, but suffers from critical implementation-level vulnerabilities, including authentication and authorization failures at its SSE gateway. Conversely, ACP's architectural flexibility, most notably its optional JWS enforcement, translates into high-impact integrity and confidentiality flaws. We contextualize these findings within current industry trends, highlighting that existing protocols remain insufficiently secure. As a path forward, we recommend a hybrid approach that combines CORAL's integrated architecture with ACP's mandatory per-message integrity guarantees, laying the groundwork for resilient, next-generation agent communications.

cs.CR↗

arXiv2025

Improving Google A2A Protocol: Protecting Sensitive Data and Mitigating Unintended Harms in Multi-Agent Systems

Googles A2A protocol provides a secure communication framework for AI agents but demonstrates critical limitations when handling highly sensitive information such as payment credentials and identity documents. These gaps increase the risk of unintended harms, including unauthorized disclosure, privilege escalation, and misuse of private data in generative multi-agent environments. In this paper, we identify key weaknesses of A2A: insufficient token lifetime control, lack of strong customer authentication, overbroad access scopes, and missing consent flows. We propose protocol-level enhancements grounded in a structured threat model for semi-trusted multi-agent systems. Our refinements introduce explicit consent orchestration, ephemeral scoped tokens, and direct user-to-service data channels to minimize exposure across time, context, and topology. Empirical evaluation using adversarial prompt injection tests shows that the enhanced protocol substantially reduces sensitive data leakage while maintaining low communication latency. Comparative analysis highlights the advantages of our approach over both the original A2A specification and related academic proposals. These contributions establish a practical path for evolving A2A into a privacy-preserving framework that mitigates unintended harms in multi-agent generative AI systems.

cs.CR↗

arXiv2024

Cloudy with a Chance of Anomalies: Dynamic Graph Neural Network for Early Detection of Cloud Services' User Anomalies

Ensuring the security of cloud environments is imperative for sustaining organizational growth and operational efficiency. As the ubiquity of cloud services continues to rise, the inevitability of cyber threats underscores the importance of preemptive detection. This paper introduces a pioneering time-based embedding approach for Cloud Services Graph-based Anomaly Detection (CS-GAD), utilizing a Graph Neural Network (GNN) to discern anomalous user behavior during interactions with cloud services. Our method employs a dynamic tripartite graph representation to encapsulate the evolving interactions among cloud services, users, and their activities over time. Leveraging GNN models in each time frame, our approach generates a graph embedding wherein each user is assigned a score based on their historical activity, facilitating the identification of unusual behavior. Results demonstrate a notable reduction in false positive rates (2-9%) compared to prevailing methods, coupled with a commendable true positive rate (100%). The contributions of this work encompass early detection capabilities, a low false positive rate, an innovative tripartite graph representation incorporating action types, the introduction of a new cloud services dataset featuring various user attacks, and an open-source implementation for community collaboration in advancing cloud service security.

cs.NI↗

arXiv2024

Enhancing Encrypted Internet Traffic Classification Through Advanced Data Augmentation Techniques

The increasing popularity of online services has made Internet Traffic Classification a critical field of study. However, the rapid development of internet protocols and encryption limits usable data availability. This paper addresses the challenges of classifying encrypted internet traffic, focusing on the scarcity of open-source datasets and limitations of existing ones. We propose two Data Augmentation (DA) techniques to synthetically generate data based on real samples: Average augmentation and MTU augmentation. Both augmentations are aimed to improve the performance of the classifier, each from a different perspective: The Average augmentation aims to increase dataset size by generating new synthetic samples, while the MTU augmentation enhances classifier robustness to varying Maximum Transmission Units (MTUs). Our experiments, conducted on two well-known academic datasets and a commercial dataset, demonstrate the effectiveness of these approaches in improving model performance and mitigating constraints associated with limited and homogeneous datasets. Our findings underscore the potential of data augmentation in addressing the challenges of modern internet traffic classification. Specifically, we show that our augmentation techniques significantly enhance encrypted traffic classification models. This improvement can positively impact user Quality of Experience (QoE) by more accurately classifying traffic as video streaming (e.g., YouTube) or chat (e.g., Google Chat). Additionally, it can enhance Quality of Service (QoS) for file downloading activities (e.g., Google Docs).

cs.LG↗

arXiv2024

RoBERTa-Augmented Synthesis for Detecting Malicious API Requests

Web applications and APIs face constant threats from malicious actors seeking to exploit vulnerabilities for illicit gains. To defend against these threats, it is essential to have anomaly detection systems that can identify a variety of malicious behaviors. However, a significant challenge in this area is the limited availability of training data. Existing datasets often do not provide sufficient coverage of the diverse API structures, parameter formats, and usage patterns encountered in real-world scenarios. As a result, models trained on these datasets often struggle to generalize and may fail to detect less common or emerging attack vectors. To enhance detection accuracy and robustness, it is crucial to access larger and more representative datasets that capture the true variability of API traffic. To address this, we introduce a GAN-inspired learning framework that extends limited API traffic datasets through targeted, domain-aware synthesis. Drawing on techniques from Natural Language Processing (NLP), our approach leverages Transformer-based architectures, particularly RoBERTa, to enhance the contextual representation of API requests and generate realistic synthetic samples aligned with security-specific semantics. We evaluate our framework on two benchmark datasets, CSIC 2010 and ATRDF 2023, and compare it with a previous data augmentation technique to assess the importance of domain-specific synthesis. In addition, we apply our augmented data to various anomaly detection models to evaluate its impact on classification performance. Our method achieves up to a 4.94% increase in F1 score on CSIC 2010 and up to 21.10% on ATRDF 2023. The source codes of this work are available at https://github.com/ArielCyber/GAN-API.

cs.CR↗

arXiv2024

A Classification-by-Retrieval Framework for Few-Shot Anomaly Detection to Detect API Injection Attacks

Application Programming Interface (API) Injection attacks refer to the unauthorized or malicious use of APIs, which are often exploited to gain access to sensitive data or manipulate online systems for illicit purposes. Identifying actors that deceitfully utilize an API poses a demanding problem. Although there have been notable advancements and contributions in the field of API security, there remains a significant challenge when dealing with attackers who use novel approaches that don't match the well-known payloads commonly seen in attacks. Also, attackers may exploit standard functionalities unconventionally and with objectives surpassing their intended boundaries. Thus, API security needs to be more sophisticated and dynamic than ever, with advanced computational intelligence methods, such as machine learning models that can quickly identify and respond to abnormal behavior. In response to these challenges, we propose a novel unsupervised few-shot anomaly detection framework composed of two main parts: First, we train a dedicated generic language model for API based on FastText embedding. Next, we use Approximate Nearest Neighbor search in a classification-by-retrieval approach. Our framework allows for training a fast, lightweight classification model using only a few examples of normal API requests. We evaluated the performance of our framework using the CSIC 2010 and ATRDF 2023 datasets. The results demonstrate that our framework improves API attack detection accuracy compared to the state-of-the-art (SOTA) unsupervised anomaly detection baselines.

cs.CR↗

arXiv2024

CBR -- Boosting Adaptive Classification By Retrieval of Encrypted Network Traffic with Out-of-distribution

Encrypted network traffic Classification tackles the problem from different approaches and with different goals. One of the common approaches is using Machine learning or Deep Learning-based solutions on a fixed number of classes, leading to misclassification when an unknown class is given as input. One of the solutions for handling unknown classes is to retrain the model, however, retraining models every time they become obsolete is both resource and time-consuming. Therefore, there is a growing need to allow classification models to detect and adapt to new classes dynamically, without retraining, but instead able to detect new classes using few shots learning [1]. In this paper, we introduce Adaptive Classification By Retrieval CBR, a novel approach for encrypted network traffic classification. Our new approach is based on an ANN-based method, which allows us to effectively identify new and existing classes without retraining the model. The novel approach is simple, yet effective and achieved similar results to RF with up to 5% difference (usually less than that) in the classification tasks while having a slight decrease in the case of new samples (from new classes) without retraining. To summarize, the new method is a real-time classification, which can classify new classes without retraining. Furthermore, our solution can be used as a complementary solution alongside RF or any other machine/deep learning classification method, as an aggregated solution.

cs.LG↗

arXiv2023

Open Image Content Disarm And Reconstruction

With the advance in malware technology, attackers create new ways to hide their malicious code from antivirus services. One way to obfuscate an attack is to use common files as cover to hide the malicious scripts, so the malware will look like a legitimate file. Although cutting-edge Artificial Intelligence and content signature exist, evasive malware successfully bypasses next-generation malware detection using advanced methods like steganography. Some of the files commonly used to hide malware are image files (e.g., JPEG). In addition, some malware use steganography to hide malicious scripts or sensitive data in images. Steganography in images is difficult to detect even with specialized tools. Image-based attacks try to attack the user's device using malicious payloads or utilize image steganography to hide sensitive data inside legitimate images and leak it outside the user's device. Therefore in this paper, we present a novel Image Content Disarm and Reconstruction (ICDR). Our ICDR system removes potential malware, with a zero trust approach, while maintaining high image quality and file usability. By extracting the image data, removing it from the rest of the file, and manipulating the image pixels, it is possible to disable or remove the hidden malware inside the file.

cs.CR↗

arXiv2022

Open-Source Framework for Encrypted Internet and Malicious Traffic Classification

Internet traffic classification plays a key role in network visibility, Quality of Services (QoS), intrusion detection, Quality of Experience (QoE) and traffic-trend analyses. In order to improve privacy, integrity, confidentiality, and protocol obfuscation, the current traffic is based on encryption protocols, e.g., SSL/TLS. With the increased use of Machine-Learning (ML) and Deep-Learning (DL) models in the literature, comparison between different models and methods has become cumbersome and difficult due to a lack of a standardized framework. In this paper, we propose an open-source framework, named OSF-EIMTC, which can provide the full pipeline of the learning process. From the well-known datasets to extracting new and well-known features, it provides implementations of well-known ML and DL models (from the traffic classification literature) as well as evaluations. Such a framework can facilitate research in traffic classification domains, so that it will be more repeatable, reproducible, easier to execute, and will allow a more accurate comparison of well-known and novel features and models. As part of our framework evaluation, we demonstrate a variety of cases where the framework can be of use, utilizing multiple datasets, models, and feature sets. We show analyses of publicly available datasets and invite the community to participate in our open challenges using the OSF-EIMTC.

cs.LG↗

arXiv2022

When a RF Beats a CNN and GRU, Together -- A Comparison of Deep Learning and Classical Machine Learning Approaches for Encrypted Malware Traffic Classification

Internet traffic classification is widely used to facilitate network management. It plays a crucial role in Quality of Services (QoS), Quality of Experience (QoE), network visibility, intrusion detection, and traffic trend analyses. While there is no theoretical guarantee that deep learning (DL)-based solutions perform better than classic machine learning (ML)-based ones, DL-based models have become the common default. This paper compares well-known DL-based and ML-based models and shows that in the case of malicious traffic classification, state-of-the-art DL-based solutions do not necessarily outperform the classical ML-based ones. We exemplify this finding using two well-known datasets for a varied set of tasks, such as: malware detection, malware family classification, detection of zero-day attacks, and classification of an iteratively growing dataset. Note that, it is not feasible to evaluate all possible models to make a concrete statement, thus, the above finding is not a recommendation to avoid DL-based models, but rather empirical proof that in some cases, there are more simplistic solutions, that may perform even better.

cs.CR↗

arXiv2022

Problem-Space Evasion Attacks in the Android OS: a Survey

Android is the most popular OS worldwide. Therefore, it is a target for various kinds of malware. As a countermeasure, the security community works day and night to develop appropriate Android malware detection systems, with ML-based or DL-based systems considered as some of the most common types. Against these detection systems, intelligent adversaries develop a wide set of evasion attacks, in which an attacker slightly modifies a malware sample to evade its target detection system. In this survey, we address problem-space evasion attacks in the Android OS, where attackers manipulate actual APKs, rather than their extracted feature vector. We aim to explore this kind of attacks, frequently overlooked by the research community due to a lack of knowledge of the Android domain, or due to focusing on general mathematical evasion attacks - i.e., feature-space evasion attacks. We discuss the different aspects of problem-space evasion attacks, using a new taxonomy, which focuses on key ingredients of each problem-space attack, such as the attacker model, the attacker's mode of operation, and the functional assessment of post-attack applications.

cs.CR↗

arXiv2022

Do You Think You Can Hold Me? The Real Challenge of Problem-Space Evasion Attacks

Android malware is a spreading disease in the virtual world. Anti-virus and detection systems continuously undergo patches and updates to defend against these threats. Most of the latest approaches in malware detection use Machine Learning (ML). Against the robustifying effort of detection systems, raise the \emph{evasion attacks}, where an adversary changes its targeted samples so that they are misclassified as benign. This paper considers two kinds of evasion attacks: feature-space and problem-space. \emph{Feature-space} attacks consider an adversary who manipulates ML features to evade the correct classification while minimizing or constraining the total manipulations. \textit{Problem-space} attacks refer to evasion attacks that change the actual sample. Specifically, this paper analyzes the gap between these two types in the Android malware domain. The gap between the two types of evasion attacks is examined via the retraining process of classifiers using each one of the evasion attack types. The experiments show that the gap between these two types of retrained classifiers is dramatic and may increase to 96\%. Retrained classifiers of feature-space evasion attacks have been found to be either less effective or completely ineffective against problem-space evasion attacks. Additionally, exploration of different problem-space evasion attacks shows that retraining of one problem-space evasion attack may be effective against other problem-space evasion attacks.

cs.CR↗

arXiv2022

MaMaDroid2.0 -- The Holes of Control Flow Graphs

Android malware is a continuously expanding threat to billions of mobile users around the globe. Detection systems are updated constantly to address these threats. However, a backlash takes the form of evasion attacks, in which an adversary changes malicious samples such that those samples will be misclassified as benign. This paper fully inspects a well-known Android malware detection system, MaMaDroid, which analyzes the control flow graph of the application. Changes to the portion of benign samples in the train set and models are considered to see their effect on the classifier. The changes in the ratio between benign and malicious samples have a clear effect on each one of the models, resulting in a decrease of more than 40% in their detection rate. Moreover, adopted ML models are implemented as well, including 5-NN, Decision Tree, and Adaboost. Exploration of the six models reveals a typical behavior in different cases, of tree-based models and distance-based models. Moreover, three novel attacks that manipulate the CFG and their detection rates are described for each one of the targeted models. The attacks decrease the detection rate of most of the models to 0%, with regards to different ratios of benign to malicious apps. As a result, a new version of MaMaDroid is engineered. This model fuses the CFG of the app and static analysis of features of the app. This improved model is proved to be robust against evasion attacks targeting both CFG-based models and static analysis models, achieving a detection rate of more than 90% against each one of the attacks.

cs.CR↗

arXiv2020

Robust Machine Learning for Encrypted Traffic Classification

Desktops and laptops can be maliciously exploited to violate privacy. In this paper, we consider the daily battle between the passive attacker who is targeting a specific user against a user that may be adversarial opponent. In this scenario, while the attacker tries to choose the best vector attack by surreptitiously monitoring the victims encrypted network traffic in order to identify users parameters such as the Operating System (OS), browser and apps. The user may use tools such as a Virtual Private Network (VPN) or even change protocols parameters to protect his/her privacy. We provide a large dataset of more than 20,000 examples for this task. We run a comprehensive set of experiments, that achieves high (above 85) classification accuracy, robustness and resilience to changes of features as a function of different network conditions at test time. We also show the effect of a small training set on the accuracy.

cs.CR↗

arXiv2020

Less is More: Robust and Novel Features for Malicious Domain Detection

Malicious domains are increasingly common and pose a severe cybersecurity threat. Specifically, many types of current cyber attacks use URLs for attack communications (e.g., C\&C, phishing, and spear-phishing). Despite the continuous progress in detecting these attacks, many alarming problems remain open, such as the weak spots of the defense mechanisms. Since machine learning has become one of the most prominent methods of malware detection, A robust feature selection mechanism is proposed that results in malicious domain detection models that are resistant to evasion attacks. This mechanism exhibits high performance based on empirical data. This paper makes two main contributions: First, it provides an analysis of robust feature selection based on widely used features in the literature. Note that even though the feature set dimensional space is reduced by half (from nine to four features), the performance of the classifier is still improved (an increase in the model's F1-score from 92.92\% to 95.81\%). Second, it introduces novel features that are robust to the adversary's manipulation. Based on an extensive evaluation of the different feature sets and commonly used classification models, this paper shows that models that are based on robust features are resistant to malicious perturbations, and at the same time useful for classifying non-manipulated data.

cs.CR↗

arXiv2020

When the Guard failed the Droid: A case study of Android malware

Android malware is a persistent threat to billions of users around the world. As a countermeasure, Android malware detection systems are occasionally implemented. However, these systems are often vulnerable to \emph{evasion attacks}, in which an adversary manipulates malicious instances so that they are misidentified as benign. In this paper, we launch various innovative evasion attacks against several Android malware detection systems. The vulnerability inherent to all of these systems is that they are part of Androguard~\cite{desnos2011androguard}, a popular open source library used in Android malware detection systems. Some of the detection systems decrease to a 0\% detection rate after the attack. Therefore, the use of open source libraries in malware detection systems calls for caution. In addition, we present a novel evaluation scheme for evasion attack generation that exploits the weak spots of known Android malware detection systems. In so doing, we evaluate the functionality and maliciousness of the manipulated instances created by our evasion attacks. We found variations in both the maliciousness and functionality tests of our manipulated apps. We show that non-functional apps, while considered malicious, do not threaten users and are thus useless from an attacker's point of view. We conclude that evasion attacks must be assessed for both functionality and maliciousness to evaluate their impact, a step which is far from commonplace today.

cs.CR↗