Alsharif Abuadbba

Ngoc Duy Pham, Alsharif Abuadbba, Yansong Gao et al.

Split learning (SL) enables data privacy preservation by allowing clients to collaboratively train a deep learning model with the server without sharing raw data. However, SL still has limitations such as potential data privacy leakage and high computation at clients. In this study, we propose to binarize the SL local layers for faster computation (up to 17.5 times less forward-propagation time in both training and inference phases on mobile devices) and reduced memory usage (up to 32 times less memory and bandwidth requirements). More importantly, the binarized SL (B-SL) model can reduce privacy leakage from SL smashed data with merely a small degradation in model accuracy. To further enhance the privacy preservation, we also propose two novel approaches: 1) training with additional local leak loss and 2) applying differential privacy, which could be integrated separately or concurrently into the B-SL model. Experimental results with different datasets have affirmed the advantages of the B-SL models compared with several benchmark models. The effectiveness of B-SL models against feature-space hijacking attack (FSHA) is also illustrated. Our results have demonstrated B-SL models are promising for lightweight IoT/mobile applications with high privacy-preservation requirements such as mobile healthcare applications.

Alsharif Abuadbba, Shuo Wang, Mahathir Almashor et al.

Web phishing remains a serious cyber threat responsible for most data breaches. Machine Learning (ML)-based anti-phishing detectors are seen as an effective countermeasure, and are increasingly adopted by web-browsers and software products. However, with an average of 10K phishing links reported per hour to platforms such as PhishTank and VirusTotal (VT), the deficiencies of such ML-based solutions are laid bare. We first explore how phishing sites bypass ML-based detection with a deep dive into 13K phishing pages targeting major brands such as Facebook. Results show successful evasion is caused by: (1) use of benign services to obscure phishing URLs; (2) high similarity between the HTML structures of phishing and benign pages; (3) hiding the ultimate phishing content within Javascript and running such scripts only on the client; (4) looking beyond typical credentials and credit cards for new content such as IDs and documents; (5) hiding phishing content until after human interaction. We attribute the root cause to the dependency of ML-based models on the vertical feature space (webpage content). These solutions rely only on what phishers present within the page itself. Thus, we propose Anti-SubtlePhish, a more resilient model based on logistic regression. The key augmentation is the inclusion of a horizontal feature space, which examines correlation variables between the final render of suspicious pages against what trusted services have recorded (e.g., PageRank). To defeat (1) and (2), we correlate information between WHOIS, PageRank, and page analytics. To combat (3), (4) and (5), we correlate features after rendering the page. Experiments with 100K phishing/benign sites show promising accuracy (98.8%). We also obtained 100% accuracy against 0-day phishing pages that were manually crafted, comparing well to the 0% recorded by VT vendors over the first four days.

Binh Le, Shahroz Tariq, Alsharif Abuadbba et al.

Recent rapid advancements in deepfake technology have allowed the creation of highly realistic fake media, such as video, image, and audio. These materials pose significant challenges to human authentication, such as impersonation, misinformation, or even a threat to national security. To keep pace with these rapid advancements, several deepfake detection algorithms have been proposed, leading to an ongoing arms race between deepfake creators and deepfake detectors. Nevertheless, these detectors are often unreliable and frequently fail to detect deepfakes. This study highlights the challenges they face in detecting deepfakes, including (1) the pre-processing pipeline of artifacts and (2) the fact that generators of new, unseen deepfake samples have not been considered when building the defense models. Our work sheds light on the need for further research and development in this field to create more robust and reliable detectors.

Hua Ma, Yinshan Li, Yansong Gao et al.

Object detection is the foundation of various critical computer-vision tasks such as segmentation, object tracking, and event detection. To train an object detector with satisfactory accuracy, a large amount of data is required. However, due to the intensive workforce involved with annotating large datasets, such a data curation task is often outsourced to a third party or relied on volunteers. This work reveals severe vulnerabilities of such data curation pipeline. We propose MACAB that crafts clean-annotated images to stealthily implant the backdoor into the object detectors trained on them even when the data curator can manually audit the images. We observe that the backdoor effect of both misclassification and the cloaking are robustly achieved in the wild when the backdoor is activated with inconspicuously natural physical triggers. Backdooring non-classification object detection with clean-annotation is challenging compared to backdooring existing image classification tasks with clean-label, owing to the complexity of having multiple objects within each frame, including victim and non-victim objects. The efficacy of the MACAB is ensured by constructively i abusing the image-scaling function used by the deep learning framework, ii incorporating the proposed adversarial clean image replica technique, and iii combining poison data selection criteria given constrained attacking budget. Extensive experiments demonstrate that MACAB exhibits more than 90% attack success rate under various real-world scenes. This includes both cloaking and misclassification backdoor effect even restricted with a small attack budget. The poisoned samples cannot be effectively identified by state-of-the-art detection techniques.The comprehensive video demo is at https://youtu.be/MA7L_LpXkp4, which is based on a poison rate of 0.14% for YOLOv4 cloaking backdoor and Faster R-CNN misclassification backdoor.

Shahroz Tariq, Alsharif Abuadbba, Kristen Moore

The metaverse has gained significant attention from various industries due to its potential to create a fully immersive and interactive virtual world. However, the integration of deepfakes in the metaverse brings serious security implications, particularly with regard to impersonation. This paper examines the security implications of deepfakes in the metaverse, specifically in the context of gaming, online meetings, and virtual offices. The paper discusses how deepfakes can be used to impersonate in gaming scenarios, how online meetings in the metaverse open the door for impersonation, and how virtual offices in the metaverse lack physical authentication, making it easier for attackers to impersonate someone. The implications of these security concerns are discussed in relation to the confidentiality, integrity, and availability (CIA) triad. The paper further explores related issues such as the darkverse, and digital cloning, as well as regulatory and privacy concerns associated with addressing security threats in the virtual world.

Huming Qiu, Hua Ma, Zhi Zhang et al.

Since Deep Learning (DL) backdoor attacks have been revealed as one of the most insidious adversarial attacks, a number of countermeasures have been developed with certain assumptions defined in their respective threat models. However, the robustness of these countermeasures is inadvertently ignored, which can introduce severe consequences, e.g., a countermeasure can be misused and result in a false implication of backdoor detection. For the first time, we critically examine the robustness of existing backdoor countermeasures with an initial focus on three influential model-inspection ones that are Neural Cleanse (S&P'19), ABS (CCS'19), and MNTD (S&P'21). Although the three countermeasures claim that they work well under their respective threat models, they have inherent unexplored non-robust cases depending on factors such as given tasks, model architectures, datasets, and defense hyper-parameter, which are \textit{not even rooted from delicate adaptive attacks}. We demonstrate how to trivially bypass them aligned with their respective threat models by simply varying aforementioned factors. Particularly, for each defense, formal proofs or empirical studies are used to reveal its two non-robust cases where it is not as robust as it claims or expects, especially the recent MNTD. This work highlights the necessity of thoroughly evaluating the robustness of backdoor countermeasures to avoid their misleading security implications in unknown non-robust cases.

Arthur Wong, Alsharif Abuadbba, Mahathir Almashor et al.

Web-based phishing accounts for over 90% of data breaches, and most web-browsers and security vendors rely on machine-learning (ML) models as mitigation. Despite this, links posted regularly on anti-phishing aggregators such as PhishTank and VirusTotal are shown to easily bypass existing detectors. Prior art suggests that automated website cloning, with light mutations, is gaining traction with attackers. This has limited exposure in current literature and leads to sub-optimal ML-based countermeasures. The work herein conducts the first empirical study that compiles and evaluates a variety of state-of-the-art cloning techniques in wide circulation. We collected 13,394 samples and found 8,566 confirmed phishing pages targeting 4 popular websites using 7 distinct cloning mechanisms. These samples were replicated with malicious code removed within a controlled platform fortified with precautions that prevent accidental access. We then reported our sites to VirusTotal and other platforms, with regular polling of results for 7 days, to ascertain the efficacy of each cloning technique. Results show that no security vendor detected our clones, proving the urgent need for more effective detectors. Finally, we posit 4 recommendations to aid web developers and ML-based defences to alleviate the risks of cloning attacks.

Hua Ma, Shang Wang, Yansong Gao et al.

All current backdoor attacks on deep learning (DL) models fall under the category of a vertical class backdoor (VCB) -- class-dependent. In VCB attacks, any sample from a class activates the implanted backdoor when the secret trigger is present. Existing defense strategies overwhelmingly focus on countering VCB attacks, especially those that are source-class-agnostic. This narrow focus neglects the potential threat of other simpler yet general backdoor types, leading to false security implications. This study introduces a new, simple, and general type of backdoor attack coined as the horizontal class backdoor (HCB) that trivially breaches the class dependence characteristic of the VCB, bringing a fresh perspective to the community. HCB is now activated when the trigger is presented together with an innocuous feature, regardless of class. For example, the facial recognition model misclassifies a person who wears sunglasses with a smiling innocuous feature into the targeted person, such as an administrator, regardless of which person. The key is that these innocuous features are horizontally shared among classes but are only exhibited by partial samples per class. Extensive experiments on attacking performance across various tasks, including MNIST, facial recognition, traffic sign recognition, object detection, and medical diagnosis, confirm the high efficiency and effectiveness of the HCB. We rigorously evaluated the evasiveness of the HCB against a series of eleven representative countermeasures, including Fine-Pruning (RAID 18'), STRIP (ACSAC 19'), Neural Cleanse (Oakland 19'), ABS (CCS 19'), Februus (ACSAC 20'), NAD (ICLR 21'), MNTD (Oakland 21'), SCAn (USENIX SEC 21'), MOTH (Oakland 22'), Beatrix (NDSS 23'), and MM-BD (Oakland 24'). None of these countermeasures prove robustness, even when employing a simplistic trigger, such as a small and static white-square patch.

Amir Kashapov, Tingmin Wu, Alsharif Abuadbba et al.

Cyber-phishing attacks recently became more precise, targeted, and tailored by training data to activate only in the presence of specific information or cues. They are adaptable to a much greater extent than traditional phishing detection. Hence, automated detection systems cannot always be 100% accurate, increasing the uncertainty around expected behavior when faced with a potential phishing email. On the other hand, human-centric defence approaches focus extensively on user training but face the difficulty of keeping users up to date with continuously emerging patterns. Therefore, advances in analyzing the content of an email in novel ways along with summarizing the most pertinent content to the recipients of emails is a prospective gateway to furthering how to combat these threats. Addressing this gap, this work leverages transformer-based machine learning to (i) analyze prospective psychological triggers, to (ii) detect possible malicious intent, and (iii) create representative summaries of emails. We then amalgamate this information and present it to the user to allow them to (i) easily decide whether the email is "phishy" and (ii) self-learn advanced malicious patterns.

Seonhye Park, Alsharif Abuadbba, Shuo Wang et al.

Training deep neural networks (DNNs) requires large datasets and powerful computing resources, which has led some owners to restrict redistribution without permission. Watermarking techniques that embed confidential data into DNNs have been used to protect ownership, but these can degrade model performance and are vulnerable to watermark removal attacks. Recently, DeepJudge was introduced as an alternative approach to measuring the similarity between a suspect and a victim model. While DeepJudge shows promise in addressing the shortcomings of watermarking, it primarily addresses situations where the suspect model copies the victim's architecture. In this study, we introduce DeepTaster, a novel DNN fingerprinting technique, to address scenarios where a victim's data is unlawfully used to build a suspect model. DeepTaster can effectively identify such DNN model theft attacks, even when the suspect model's architecture deviates from the victim's. To accomplish this, DeepTaster generates adversarial images with perturbations, transforms them into the Fourier frequency domain, and uses these transformed images to identify the dataset used in a suspect model. The underlying premise is that adversarial images can capture the unique characteristics of DNNs built with a specific dataset. To demonstrate the effectiveness of DeepTaster, we evaluated the effectiveness of DeepTaster by assessing its detection accuracy on three datasets (CIFAR10, MNIST, and Tiny-ImageNet) across three model architectures (ResNet18, VGG16, and DenseNet161). We conducted experiments under various attack scenarios, including transfer learning, pruning, fine-tuning, and data augmentation. Specifically, in the Multi-Architecture Attack scenario, DeepTaster was able to identify all the stolen cases across all datasets, while DeepJudge failed to detect any of the cases.

Mariya Shmalko, Alsharif Abuadbba, Raj Gaire et al.

Email phishing has become more prevalent and grows more sophisticated over time. To combat this rise, many machine learning (ML) algorithms for detecting phishing emails have been developed. However, due to the limited email data sets on which these algorithms train, they are not adept at recognising varied attacks and, thus, suffer from concept drift; attackers can introduce small changes in the statistical characteristics of their emails or websites to successfully bypass detection. Over time, a gap develops between the reported accuracy from literature and the algorithm's actual effectiveness in the real world. This realises itself in frequent false positive and false negative classifications. To this end, we propose a multidimensional risk assessment of emails to reduce the feasibility of an attacker adapting their email and avoiding detection. This horizontal approach to email phishing detection profiles an incoming email on its main features. We develop a risk assessment framework that includes three models which analyse an email's (1) threat level, (2) cognitive manipulation, and (3) email type, which we combine to return the final risk assessment score. The Profiler does not require large data sets to train on to be effective and its analysis of varied email features reduces the impact of concept drift. Our Profiler can be used in conjunction with ML approaches, to reduce their misclassifications or as a labeller for large email data sets in the training stage. We evaluate the efficacy of the Profiler against a machine learning ensemble using state-of-the-art ML algorithms on a data set of 9000 legitimate and 900 phishing emails from a large Australian research organisation. Our results indicate that the Profiler's mitigates the impact of concept drift, and delivers 30% less false positive and 25% less false negative email classifications over the ML ensemble's approach.

Hossein Rahimpour, Joe Tusek, Alsharif Abuadbba et al.

The rise of cyber threats on critical infrastructure and its potential for devastating consequences, has significantly increased. The dependency of new power grid technology on information, data analytic and communication systems make the entire electricity network vulnerable to cyber threats. Power transformers play a critical role within the power grid and are now commonly enhanced through factory add-ons or intelligent monitoring systems added later to improve the condition monitoring of critical and long lead time assets such as transformers. However, the increased connectivity of those power transformers opens the door to more cyber attacks. Therefore, the need to detect and prevent cyber threats is becoming critical. The first step towards that would be a deeper understanding of the potential cyber-attacks landscape against power transformers. Much of the existing literature pays attention to smart equipment within electricity distribution networks, and most methods proposed are based on model-based detection algorithms. Moreover, only a few of these works address the security vulnerabilities of power elements, especially transformers within the transmission network. To the best of our knowledge, there is no study in the literature that systematically investigate the cybersecurity challenges against the newly emerged smart transformers. This paper addresses this shortcoming by exploring the vulnerabilities and the attack vectors of power transformers within electricity networks, the possible attack scenarios and the risks associated with these attacks.

Nguyen Linh Bao Nguyen, Wanlun Ma, Viet Vo et al.

Retrieval-augmented generation (RAG) has become central to large language model (LLM) deployments, grounding responses in enterprise or proprietary data to reduce hallucinations. However, this design introduces a new privacy risk: model outputs may signal the presence of specific documents in the retrieval corpus, enabling membership inference attacks (MIAs) that leak sensitive information. Existing MIAs are feasible, but they often rely on easily detected templated queries or require many non-templated yet costly and repetitive queries, limiting practicality. We ask: Can an adversary launch a limited-budget, surrogate-free, stealthy, and defense-agnostic membership inference attack using non-templated queries? We present MEntA (Membership Entailment Attack), a query-efficient MIA that leverages natural-language entailment to maximize information gained per query. By asking low-cost, broad, information-seeking questions and measuring entailment between model responses and candidate documents, MEntA eliminates the need for costly shadow models and large query budgets. Across NFCorpus, SCIDOCS, and TREC-COVID, MEntA achieves up to 0.991 AUC with only 5 queries, outperforming prior methods by 0.20 to 0.50 AUC under equivalent conditions. It remains effective under state-of-the-art (SOTA) RAG defenses, while current detectors either miss MEntA or flag benign queries at high rates. Regarding cost, MEntA reduces total attack cost by up to 65 $\times$ lower compared to SOTA attacks under the same attack setting. Our findings expose the feasibility of realistic, low-cost privacy leakage in RAG systems and highlight the urgent need for privacy-aware retrieval and defense mechanisms.

William Guanting Li, Alsharif Abuadbba, Kristen Moore et al.

Penetration testing is essential to securing modern web infrastructures, yet traditional manual methods struggle to keep pace with their scale and complexity. Large Language Models (LLMs) offer new opportunities for automating these tasks, but existing approaches face two persistent challenges: hallucination of technical entities and insufficient long-term contextual memory. To address these issues, we present APT-Agent, a fully automated LLM-driven penetration testing framework that systematically orchestrates reconnaissance, exploitation, and exfiltration. APT-Agent introduces a hybrid rectification module to recover hallucinated commands and a command-specific memory architecture to preserve operational context across multi-step attack sequences. We evaluate our APT-Agent on Metasploitable 2 against seven vulnerable services spanning web, database, and network protocols. APT-Agent achieves an 84.29% end-to-end exploitation success rate, compared to 48.57% (Script Kiddie) and 18.57% (PentestGPT) under matched conditions. By reducing cognitive burden and minimizing reliance on human intervention, APT-Agent represents a step toward scalable, reliable, and cognitively efficient automation for penetration testing.

Alsharif Abuadbba, Nazatul Sultan, Surya Nepal et al.

AI is moving from domain-specific autonomy in closed, predictable settings to large-language-model-driven agents that plan and act in open, cross-organizational environments. As a result, the cybersecurity risk landscape is changing in fundamental ways. Agentic AI systems can plan, act, collaborate, and persist over time, functioning as participants in complex socio-technical ecosystems rather than as isolated software components. Although recent work has strengthened defenses against model and pipeline level vulnerabilities such as prompt injection, data poisoning, and tool misuse, these system centric approaches may fail to capture risks that arise from autonomy, interaction, and emergent behavior. This article introduces the 4C Framework for multi-agent AI security, inspired by societal governance. It organizes agentic risks across four interdependent dimensions: Core (system, infrastructure, and environmental integrity), Connection (communication, coordination, and trust), Cognition (belief, goal, and reasoning integrity), and Compliance (ethical, legal, and institutional governance). By shifting AI security from a narrow focus on system-centric protection to the broader preservation of behavioral integrity and intent, the framework complements existing AI security strategies and offers a principled foundation for building agentic AI systems that are trustworthy, governable, and aligned with human values.

Bushra Sabir, Yansong Gao, Alsharif Abuadbba et al.

Transformer-based text classifiers such as BERT, RoBERTa, T5, and GPT have shown strong performance in natural language processing tasks but remain vulnerable to adversarial examples. These vulnerabilities raise significant security concerns, as small input perturbations can cause severe misclassifications. Existing robustness methods often require heavy computation or lack interpretability. This paper presents a unified framework called Explainability-driven Detection, Identification, and Transformation (EDIT) to strengthen inference-time defenses. EDIT integrates explainability tools, including attention maps and integrated gradients, with frequency-based features to automatically detect and identify adversarial perturbations while offering insight into model behavior. After detection, EDIT refines adversarial inputs using an optimal transformation process that leverages pre-trained embeddings and model feedback to replace corrupted tokens. To enhance security assurance, EDIT incorporates automated alerting mechanisms that involve human analysts when necessary. Beyond static defenses, EDIT also provides adaptive resilience by enforcing internal feature similarity and transforming inputs, thereby disrupting the attackers optimization process and limiting the effectiveness of adaptive adversarial attacks. Experiments using BERT and RoBERTa on IMDB, YELP, AGNEWS, and SST2 datasets against seven word substitution attacks demonstrate that EDIT achieves an average Fscore of 89.69 percent and balanced accuracy of 89.70 percent. Compared to four state-of-the-art defenses, EDIT improves balanced accuracy by 1.22 times and F1-score by 1.33 times while being 83 times faster in feature extraction. The framework provides robust, interpretable, and efficient protection against both standard, zero-day, and adaptive adversarial threats in text classification models.

Zahra Mousavi, Chadni Islam, Kristen Moore et al.

The increasing trend of using Large Language Models (LLMs) for code generation raises the question of their capability to generate trustworthy code. While many researchers are exploring the utility of code generation for uncovering software vulnerabilities, one crucial but often overlooked aspect is the security Application Programming Interfaces (APIs). APIs play an integral role in upholding software security, yet effectively integrating security APIs presents substantial challenges. This leads to inadvertent misuse by developers, thereby exposing software to vulnerabilities. To overcome these challenges, developers may seek assistance from LLMs. In this paper, we systematically assess ChatGPT's trustworthiness in code generation for security API use cases in Java. To conduct a thorough evaluation, we compile an extensive collection of 48 programming tasks for 5 widely used security APIs. We employ both automated and manual approaches to effectively detect security API misuse in the code generated by ChatGPT for these tasks. Our findings are concerning: around 70% of the code instances across 30 attempts per task contain security API misuse, with 20 distinct misuse types identified. Moreover, for roughly half of the tasks, this rate reaches 100%, indicating that there is a long way to go before developers can rely on ChatGPT to securely implement security API code.

Binh M. Le, Jiwon Kim, Simon S. Woo et al.

Deepfakes have rapidly emerged as a serious threat to society due to their ease of creation and dissemination, triggering the accelerated development of detection technologies. However, many existing detectors rely on labgenerated datasets for validation, which may not prepare them for novel, real-world deepfakes. This paper extensively reviews and analyzes state-of-the-art deepfake detectors, evaluating them against several critical criteria. These criteria categorize detectors into 4 high-level groups and 13 finegrained sub-groups, aligned with a unified conceptual framework we propose. This classification offers practical insights into the factors affecting detector efficacy. We evaluate the generalizability of 16 leading detectors across comprehensive attack scenarios, including black-box, white-box, and graybox settings. Our systematized analysis and experiments provide a deeper understanding of deepfake detectors and their generalizability, paving the way for future research and the development of more proactive defenses against deepfakes.

Tingmin Wu, Shuiqiao Yang, Shigang Liu et al.

Threat modeling is a crucial component of cybersecurity, particularly for industries such as banking, where the security of financial data is paramount. Traditional threat modeling approaches require expert intervention and manual effort, often leading to inefficiencies and human error. The advent of Large Language Models (LLMs) offers a promising avenue for automating these processes, enhancing both efficiency and efficacy. However, this transition is not straightforward due to three main challenges: (1) the lack of publicly available, domain-specific datasets, (2) the need for tailored models to handle complex banking system architectures, and (3) the requirement for real-time, adaptive mitigation strategies that align with compliance standards like NIST 800-53. In this paper, we introduce ThreatModeling-LLM, a novel and adaptable framework that automates threat modeling for banking systems using LLMs. ThreatModeling-LLM operates in three stages: 1) dataset creation, 2) prompt engineering and 3) model fine-tuning. We first generate a benchmark dataset using Microsoft Threat Modeling Tool (TMT). Then, we apply Chain of Thought (CoT) and Optimization by PROmpting (OPRO) on the pre-trained LLMs to optimize the initial prompt. Lastly, we fine-tune the LLM using Low-Rank Adaptation (LoRA) based on the benchmark dataset and the optimized prompt to improve the threat identification and mitigation generation capabilities of pre-trained LLMs.

Erik Buchholz, Alsharif Abuadbba, Shuo Wang et al.

While location trajectories represent a valuable data source for analyses and location-based services, they can reveal sensitive information, such as political and religious preferences. Differentially private publication mechanisms have been proposed to allow for analyses under rigorous privacy guarantees. However, the traditional protection schemes suffer from a limiting privacy-utility trade-off and are vulnerable to correlation and reconstruction attacks. Synthetic trajectory data generation and release represent a promising alternative to protection algorithms. While initial proposals achieve remarkable utility, they fail to provide rigorous privacy guarantees. This paper proposes a framework for designing a privacy-preserving trajectory publication approach by defining five design goals, particularly stressing the importance of choosing an appropriate Unit of Privacy. Based on this framework, we briefly discuss the existing trajectory protection approaches, emphasising their shortcomings. This work focuses on the systematisation of the state-of-the-art generative models for trajectories in the context of the proposed framework. We find that no existing solution satisfies all requirements. Thus, we perform an experimental study evaluating the applicability of six sequential generative models to the trajectory domain. Finally, we conclude that a generative trajectory model providing semantic guarantees remains an open research question and propose concrete next steps for future research.

Xinguo Feng, Zhongkui Ma, Zihan Wang et al.

The gradient inversion attack has been demonstrated as a significant privacy threat to federated learning (FL), particularly in continuous domains such as vision models. In contrast, it is often considered less effective or highly dependent on impractical training settings when applied to language models, due to the challenges posed by the discrete nature of tokens in text data. As a result, its potential privacy threats remain largely underestimated, despite FL being an emerging training method for language models. In this work, we propose a domain-specific gradient inversion attack named Grab (gradient inversion with hybrid optimization). Grab features two alternating optimization processes to address the challenges caused by practical training settings, including a simultaneous optimization on dropout masks between layers for improved token recovery and a discrete optimization for effective token sequencing. Grab can recover a significant portion (up to 92.9% recovery rate) of the private training data, outperforming the attack strategy of utilizing discrete optimization with an auxiliary model by notable improvements of up to 28.9% recovery rate in benchmark settings and 48.5% recovery rate in practical settings. Grab provides a valuable step forward in understanding this privacy threat in the emerging FL training mode of language models.

Shahroz Tariq, David Nguyen, M. A. P. Chamikara et al.

The growing sophistication of deepfakes presents substantial challenges to the integrity of media and the preservation of public trust. Concurrently, vision-language models (VLMs), large language models enhanced with visual reasoning capabilities, have emerged as promising tools across various domains, sparking interest in their applicability to deepfake detection. This study conducts a structured zero-shot evaluation of four prominent VLMs: ChatGPT, Claude, Gemini, and Grok, focusing on three primary deepfake types: faceswap, reenactment, and synthetic generation. Leveraging a meticulously assembled benchmark comprising authentic and manipulated images from diverse sources, we evaluate each model's classification accuracy and reasoning depth. Our analysis indicates that while VLMs can produce coherent explanations and detect surface-level anomalies, they are not yet dependable as standalone detection systems. We highlight critical failure modes, such as an overemphasis on stylistic elements and vulnerability to misleading visual patterns like vintage aesthetics. Nevertheless, VLMs exhibit strengths in interpretability and contextual analysis, suggesting their potential to augment human expertise in forensic workflows. These insights imply that although general-purpose models currently lack the reliability needed for autonomous deepfake detection, they hold promise as integral components in hybrid or human-in-the-loop detection frameworks.

Tom Roth, Inigo Jauregi Unanue, Alsharif Abuadbba et al.

Text classifiers are vulnerable to adversarial examples -- correctly-classified examples that are deliberately transformed to be misclassified while satisfying acceptability constraints. The conventional approach to finding adversarial examples is to define and solve a combinatorial optimisation problem over a space of allowable transformations. While effective, this approach is slow and limited by the choice of transformations. An alternate approach is to directly generate adversarial examples by fine-tuning a pre-trained language model, as is commonly done for other text-to-text tasks. This approach promises to be much quicker and more expressive, but is relatively unexplored. For this reason, in this work we train an encoder-decoder paraphrase model to generate a diverse range of adversarial examples. For training, we adopt a reinforcement learning algorithm and propose a constraint-enforcing reward that promotes the generation of valid adversarial examples. Experimental results over two text classification datasets show that our model has achieved a higher success rate than the original paraphrase model, and overall has proved more effective than other competitive attacks. Finally, we show how key design choices impact the generated examples and discuss the strengths and weaknesses of the proposed approach.

Robert Marlin, Raja Jurdak, Alsharif Abuadbba et al.

By 2050, electric vehicles (EVs) are projected to account for 70% of global vehicle sales. While EVs provide environmental benefits, they also pose challenges for energy generation, grid infrastructure, and data privacy. Current research on EV routing and charge management often overlooks privacy when predicting energy demands, leaving sensitive mobility data vulnerable. To address this, we developed a Federated Learning Transformer Network (FLTN) to predict EVs' next charge location with enhanced privacy measures. Each EV operates as a client, training an onboard FLTN model that shares only model weights, not raw data with a community-based Distributed Energy Resource Management System (DERMS), which aggregates them into a community global model. To further enhance privacy, non-transitory EVs use peer-to-peer weight sharing and augmentation within their community, obfuscating individual contributions and improving model accuracy. Community DERMS global model weights are then redistributed to EVs for continuous training. Our FLTN approach achieved up to 92% accuracy while preserving data privacy, compared to our baseline centralised model, which achieved 98% accuracy with no data privacy. Simulations conducted across diverse charge levels confirm the FLTN's ability to forecast energy demands over extended periods. We present a privacy-focused solution for forecasting EV charge location prediction, effectively mitigating data leakage risks.

Nguyen Linh Bao Nguyen, Alsharif Abuadbba, Kristen Moore et al.

The rapid advancement of generative models has enabled the creation of increasingly stealthy synthetic voices, commonly referred to as audio deepfakes. A recent technique, FOICE [USENIX'24], demonstrates a particularly alarming capability: generating a victim's voice from a single facial image, without requiring any voice sample. By exploiting correlations between facial and vocal features, FOICE produces synthetic voices realistic enough to bypass industry-standard authentication systems, including WeChat Voiceprint and Microsoft Azure. This raises serious security concerns, as facial images are far easier for adversaries to obtain than voice samples, dramatically lowering the barrier to large-scale attacks. In this work, we investigate two core research questions: (RQ1) can state-of-the-art audio deepfake detectors reliably detect FOICE-generated speech under clean and noisy conditions, and (RQ2) whether fine-tuning these detectors on FOICE data improves detection without overfitting, thereby preserving robustness to unseen voice generators such as SpeechT5. Our study makes three contributions. First, we present the first systematic evaluation of FOICE detection, showing that leading detectors consistently fail under both standard and noisy conditions. Second, we introduce targeted fine-tuning strategies that capture FOICE-specific artifacts, yielding significant accuracy improvements. Third, we assess generalization after fine-tuning, revealing trade-offs between specialization to FOICE and robustness to unseen synthesis pipelines. These findings expose fundamental weaknesses in today's defenses and motivate new architectures and training protocols for next-generation audio deepfake detection.

Fanzhen Liu, Alsharif Abuadbba, Kristen Moore et al.

In an era where misinformation spreads freely, fact-checking (FC) plays a crucial role in verifying claims and promoting reliable information. While automated fact-checking (AFC) has advanced significantly, existing systems remain vulnerable to adversarial attacks that manipulate or generate claims, evidence, or claim-evidence pairs. These attacks can distort the truth, mislead decision-makers, and ultimately undermine the reliability of FC models. Despite growing research interest in adversarial attacks against AFC systems, a comprehensive, holistic overview of key challenges remains lacking. These challenges include understanding attack strategies, assessing the resilience of current models, and identifying ways to enhance robustness. This survey provides the first in-depth review of adversarial attacks targeting FC, categorizing existing attack methodologies and evaluating their impact on AFC systems. Additionally, we examine recent advancements in adversary-aware defenses and highlight open research questions that require further exploration. Our findings underscore the urgent need for resilient FC frameworks capable of withstanding adversarial manipulations in pursuit of preserving high verification accuracy.

Fanzhen Liu, Xiaoxiao Ma, Jian Yang et al.

Enhancing the interpretability of graph neural networks (GNNs) is crucial to ensure their safe and fair deployment. Recent work has introduced self-explainable GNNs that generate explanations as part of training, improving both faithfulness and efficiency. Some of these models, such as ProtGNN and PGIB, learn class-specific prototypes, offering a potential pathway toward class-level explanations. However, their evaluations focus solely on instance-level explanations, leaving open the question of whether these prototypes meaningfully generalize across instances of the same class. In this paper, we introduce GraphOracle, a novel self-explainable GNN framework designed to generate and evaluate class-level explanations for GNNs. Our model jointly learns a GNN classifier and a set of structured, sparse subgraphs that are discriminative for each class. We propose a novel integrated training that captures graph$\unicode{x2013}$subgraph$\unicode{x2013}$prediction dependencies efficiently and faithfully, validated through a masking-based evaluation strategy. This strategy enables us to retroactively assess whether prior methods like ProtGNN and PGIB deliver effective class-level explanations. Our results show that they do not. In contrast, GraphOracle achieves superior fidelity, explainability, and scalability across a range of graph classification tasks. We further demonstrate that GraphOracle avoids the computational bottlenecks of previous methods$\unicode{x2014}$like Monte Carlo Tree Search$\unicode{x2014}$by using entropy-regularized subgraph selection and lightweight random walk extraction, enabling faster and more scalable training. These findings position GraphOracle as a practical and principled solution for faithful class-level self-explainability in GNNs.

Erik Buchholz, Natasha Fernandes, David D. Nguyen et al.

While location trajectories offer valuable insights, they also reveal sensitive personal information. Differential Privacy (DP) offers formal protection, but achieving a favourable utility-privacy trade-off remains challenging. Recent works explore deep learning-based generative models to produce synthetic trajectories. However, current models lack formal privacy guarantees and rely on conditional information derived from real data during generation. This work investigates the utility cost of enforcing DP in such models, addressing three research questions across two datasets and eleven utility metrics. (1) We evaluate how DP-SGD, the standard DP training method for deep learning, affects the utility of state-of-the-art generative models. (2) Since DP-SGD is limited to unconditional models, we propose a novel DP mechanism for conditional generation that provides formal guarantees and assess its impact on utility. (3) We analyse how model types - Diffusion, VAE, and GAN - affect the utility-privacy trade-off. Our results show that DP-SGD significantly impacts performance, although some utility remains if the datasets is sufficiently large. The proposed DP mechanism improves training stability, particularly when combined with DP-SGD, for unstable models such as GANs and on smaller datasets. Diffusion models yield the best utility without guarantees, but with DP-SGD, GANs perform best, indicating that the best non-private model is not necessarily optimal when targeting formal guarantees. In conclusion, DP trajectory generation remains a challenging task, and formal guarantees are currently only feasible with large datasets and in constrained use cases.

Robert Marlin, Raja Jurdak, Alsharif Abuadbba

The widespread adoption of Electric Vehicles (EVs) poses critical challenges for energy providers, particularly in predicting charging time (temporal prediction), ensuring user privacy, and managing resources efficiently in mobility-driven networks. This paper introduces the Hierarchical Federated Learning Transformer Network (H-FLTN) framework to address these challenges. H-FLTN employs a three-tier hierarchical architecture comprising EVs, community Distributed Energy Resource Management Systems (DERMS), and the Energy Provider Data Centre (EPDC) to enable accurate spatio-temporal predictions of EV charging needs while preserving privacy. Temporal prediction is enhanced using Transformer-based learning, capturing complex dependencies in charging behavior. Privacy is ensured through Secure Aggregation, Additive Secret Sharing, and Peer-to-Peer (P2P) Sharing with Augmentation, which allow only secret shares of model weights to be exchanged while securing all transmissions. To improve training efficiency and resource management, H-FLTN integrates Dynamic Client Capping Mechanism (DCCM) and Client Rotation Management (CRM), ensuring that training remains both computationally and temporally efficient as the number of participating EVs increases. DCCM optimises client participation by limiting excessive computational loads, while CRM balances training contributions across epochs, preventing imbalanced participation. Our simulation results based on large-scale empirical vehicle mobility data reveal that DCCM and CRM reduce the training time complexity with increasing EVs from linear to constant. Its integration into real-world smart city infrastructure enhances energy demand forecasting, resource allocation, and grid stability, ensuring reliability and sustainability in future mobility ecosystems.

Tom Roth, Inigo Jauregi Unanue, Alsharif Abuadbba et al.

Current adversarial attack algorithms, where an adversary changes a text to fool a victim model, have been repeatedly shown to be effective against text classifiers. These attacks, however, generally assume that the victim model is monolingual and cannot be used to target multilingual victim models, a significant limitation given the increased use of these models. For this reason, in this work we propose an approach to fine-tune a multilingual paraphrase model with an adversarial objective so that it becomes able to generate effective adversarial examples against multilingual classifiers. The training objective incorporates a set of pre-trained models to ensure text quality and language consistency of the generated text. In addition, all the models are suitably connected to the generator by vocabulary-mapping matrices, allowing for full end-to-end differentiability of the overall training pipeline. The experimental validation over two multilingual datasets and five languages has shown the effectiveness of the proposed approach compared to existing baselines, particularly in terms of query efficiency. We also provide a detailed analysis of the generated attacks and discuss limitations and opportunities for future research.

Hua Ma, Yinshan Li, Yansong Gao et al.

Deep learning models have been shown to be vulnerable to recent backdoor attacks. A backdoored model behaves normally for inputs containing no attacker-secretly-chosen trigger and maliciously for inputs with the trigger. To date, backdoor attacks and countermeasures mainly focus on image classification tasks. And most of them are implemented in the digital world with digital triggers. Besides the classification tasks, object detection systems are also considered as one of the basic foundations of computer vision tasks. However, there is no investigation and understanding of the backdoor vulnerability of the object detector, even in the digital world with digital triggers. For the first time, this work demonstrates that existing object detectors are inherently susceptible to physical backdoor attacks. We use a natural T-shirt bought from a market as a trigger to enable the cloaking effect--the person bounding-box disappears in front of the object detector. We show that such a backdoor can be implanted from two exploitable attack scenarios into the object detector, which is outsourced or fine-tuned through a pretrained model. We have extensively evaluated three popular object detection algorithms: anchor-based Yolo-V3, Yolo-V4, and anchor-free CenterNet. Building upon 19 videos shot in real-world scenes, we confirm that the backdoor attack is robust against various factors: movement, distance, angle, non-rigid deformation, and lighting. Specifically, the attack success rate (ASR) in most videos is 100% or close to it, while the clean data accuracy of the backdoored model is the same as its clean counterpart. The latter implies that it is infeasible to detect the backdoor behavior merely through a validation set. The averaged ASR still remains sufficiently high to be 78% in the transfer learning attack scenarios evaluated on CenterNet. See the demo video on https://youtu.be/Q3HOF4OobbY.

Yinshan Li, Hua Ma, Zhi Zhang et al.

A backdoor deep learning (DL) model behaves normally upon clean inputs but misbehaves upon trigger inputs as the backdoor attacker desires, posing severe consequences to DL model deployments. State-of-the-art defenses are either limited to specific backdoor attacks (source-agnostic attacks) or non-user-friendly in that machine learning (ML) expertise or expensive computing resources are required. This work observes that all existing backdoor attacks have an inevitable intrinsic weakness, non-transferability, that is, a trigger input hijacks a backdoored model but cannot be effective to another model that has not been implanted with the same backdoor. With this key observation, we propose non-transferability enabled backdoor detection (NTD) to identify trigger inputs for a model-under-test (MUT) during run-time.Specifically, NTD allows a potentially backdoored MUT to predict a class for an input. In the meantime, NTD leverages a feature extractor (FE) to extract feature vectors for the input and a group of samples randomly picked from its predicted class, and then compares similarity between the input and the samples in the FE's latent space. If the similarity is low, the input is an adversarial trigger input; otherwise, benign. The FE is a free pre-trained model privately reserved from open platforms. As the FE and MUT are from different sources, the attacker is very unlikely to insert the same backdoor into both of them. Because of non-transferability, a trigger effect that does work on the MUT cannot be transferred to the FE, making NTD effective against different types of backdoor attacks. We evaluate NTD on three popular customized tasks such as face recognition, traffic sign recognition and general animal classification, results of which affirm that NDT has high effectiveness (low false acceptance rate) and usability (low false rejection rate) with low detection latency.

Hua Ma, Huming Qiu, Yansong Gao et al.

Currently, there is a burgeoning demand for deploying deep learning (DL) models on ubiquitous edge Internet of Things (IoT) devices attributed to their low latency and high privacy preservation. However, DL models are often large in size and require large-scale computation, which prevents them from being placed directly onto IoT devices, where resources are constrained and 32-bit floating-point (float-32) operations are unavailable. Commercial framework (i.e., a set of toolkits) empowered model quantization is a pragmatic solution that enables DL deployment on mobile devices and embedded systems by effortlessly post-quantizing a large high-precision model (e.g., float-32) into a small low-precision model (e.g., int-8) while retaining the model inference accuracy. However, their usability might be threatened by security vulnerabilities. This work reveals that the standard quantization toolkits can be abused to activate a backdoor. We demonstrate that a full-precision backdoored model which does not have any backdoor effect in the presence of a trigger -- as the backdoor is dormant -- can be activated by the default i) TensorFlow-Lite (TFLite) quantization, the only product-ready quantization framework to date, and ii) the beta released PyTorch Mobile framework. When each of the float-32 models is converted into an int-8 format model through the standard TFLite or Pytorch Mobile framework's post-training quantization, the backdoor is activated in the quantized model, which shows a stable attack success rate close to 100% upon inputs with the trigger, while it behaves normally upon non-trigger inputs. This work highlights that a stealthy security threat occurs when an end user utilizes the on-device post-training model quantization frameworks, informing security researchers of cross-platform overhaul of DL models post quantization even if these models pass front-end backdoor inspections.

Keelan Evans, Alsharif Abuadbba, Tingmin Wu et al.

Spear Phishing is a harmful cyber-attack facing business and individuals worldwide. Considerable research has been conducted recently into the use of Machine Learning (ML) techniques to detect spear-phishing emails. ML-based solutions may suffer from zero-day attacks; unseen attacks unaccounted for in the training data. As new attacks emerge, classifiers trained on older data are unable to detect these new varieties of attacks resulting in increasingly inaccurate predictions. Spear Phishing detection also faces scalability challenges due to the growth of the required features which is proportional to the number of the senders within a receiver mailbox. This differs from traditional phishing attacks which typically perform only a binary classification between phishing and benign emails. Therefore, we devise a possible solution to these problems, named RAIDER: Reinforcement AIded Spear Phishing DEtectoR. A reinforcement-learning based feature evaluation system that can automatically find the optimum features for detecting different types of attacks. By leveraging a reward and penalty system, RAIDER allows for autonomous features selection. RAIDER also keeps the number of features to a minimum by selecting only the significant features to represent phishing emails and detect spear-phishing attacks. After extensive evaluation of RAIDER over 11,000 emails and across 3 attack scenarios, our results suggest that using reinforcement learning to automatically identify the significant features could reduce the dimensions of the required features by 55% in comparison to existing ML-based systems. It also improves the accuracy of detecting spoofing attacks by 4% from 90% to 94%. In addition, RAIDER demonstrates reasonable detection accuracy even against a sophisticated attack named Known Sender in which spear-phishing emails greatly resemble those of the impersonated sender.

Shuo Wang, Surya Nepal, Kristen Moore et al.

The diversity and quantity of data warehouses, gathering data from distributed devices such as mobile devices, can enhance the success and robustness of machine learning algorithms. Federated learning enables distributed participants to collaboratively learn a commonly-shared model while holding data locally. However, it is also faced with expensive communication and limitations due to the heterogeneity of distributed data sources and lack of access to global data. In this paper, we investigate a practical distributed learning scenario where multiple downstream tasks (e.g., classifiers) could be efficiently learned from dynamically-updated and non-iid distributed data sources while providing local data privatization. We introduce a new distributed/collaborative learning scheme to address communication overhead via latent compression, leveraging global data while providing privatization of local data without additional cost due to encryption or perturbation. This scheme divides learning into (1) informative feature encoding, and transmitting the latent representation of local data to address communication overhead; (2) downstream tasks centralized at the server using the encoded codes gathered from each node to address computing overhead. Besides, a disentanglement strategy is applied to address the privatization of sensitive components of local data. Extensive experiments are conducted on image and speech datasets. The results demonstrate that downstream tasks on the compact latent representations with the privatization of local data can achieve comparable accuracy to centralized learning.

Tom Roth, Yansong Gao, Alsharif Abuadbba et al.

Many adversarial attacks target natural language processing systems, most of which succeed through modifying the individual tokens of a document. Despite the apparent uniqueness of each of these attacks, fundamentally they are simply a distinct configuration of four components: a goal function, allowable transformations, a search method, and constraints. In this survey, we systematically present the different components used throughout the literature, using an attack-independent framework which allows for easy comparison and categorisation of components. Our work aims to serve as a comprehensive guide for newcomers to the field and to spark targeted research into refining the individual attack components.

Alsharif Abuadbba, Hyoungshick Kim, Surya Nepal

Convolutional Neural Networks (CNNs) deployed in real-life applications such as autonomous vehicles have shown to be vulnerable to manipulation attacks, such as poisoning attacks and fine-tuning. Hence, it is essential to ensure the integrity and authenticity of CNNs because compromised models can produce incorrect outputs and behave maliciously. In this paper, we propose a self-contained tamper-proofing method, called DeepiSign, to ensure the integrity and authenticity of CNN models against such manipulation attacks. DeepiSign applies the idea of fragile invisible watermarking to securely embed a secret and its hash value into a CNN model. To verify the integrity and authenticity of the model, we retrieve the secret from the model, compute the hash value of the secret, and compare it with the embedded hash value. To minimize the effects of the embedded secret on the CNN model, we use a wavelet-based technique to transform weights into the frequency domain and embed the secret into less significant coefficients. Our theoretical analysis shows that DeepiSign can hide up to 1KB secret in each layer with minimal loss of the model's accuracy. To evaluate the security and performance of DeepiSign, we performed experiments on four pre-trained models (ResNet18, VGG16, AlexNet, and MobileNet) using three datasets (MNIST, CIFAR-10, and Imagenet) against three types of manipulation attacks (targeted input poisoning, output poisoning, and fine-tuning). The results demonstrate that DeepiSign is verifiable without degrading the classification accuracy, and robust against representative CNN manipulation attacks.

Bedeuro Kim, Alsharif Abuadbba, Yansong Gao et al.

As an essential processing step in computer vision applications, image resizing or scaling, more specifically downsampling, has to be applied before feeding a normally large image into a convolutional neural network (CNN) model because CNN models typically take small fixed-size images as inputs. However, image scaling functions could be adversarially abused to perform a newly revealed attack called image-scaling attack, which can affect a wide range of computer vision applications building upon image-scaling functions. This work presents an image-scaling attack detection framework, termed as Decamouflage. Decamouflage consists of three independent detection methods: (1) rescaling, (2) filtering/pooling, and (3) steganalysis. While each of these three methods is efficient standalone, they can work in an ensemble manner not only to improve the detection accuracy but also to harden potential adaptive attacks. Decamouflage has a pre-determined detection threshold that is generic. More precisely, as we have validated, the threshold determined from one dataset is also applicable to other different datasets. Extensive experiments show that Decamouflage achieves detection accuracy of 99.9\% and 99.8\% in the white-box (with the knowledge of attack algorithms) and the black-box (without the knowledge of attack algorithms) settings, respectively. To corroborate the efficiency of Decamouflage, we have also measured its run-time overhead on a personal PC with an i5 CPU and found that Decamouflage can detect image-scaling attacks in milliseconds. Overall, Decamouflage can accurately detect image scaling attacks in both white-box and black-box settings with acceptable run-time overhead.

Chandra Thapa, Jun Wen Tang, Alsharif Abuadbba et al.

The use of Artificial Intelligence (AI) to detect phishing emails is primarily dependent on large-scale centralized datasets, which opens it up to a myriad of privacy, trust, and legal issues. Moreover, organizations are loathed to share emails, given the risk of leakage of commercially sensitive information. So, it is uncommon to obtain sufficient emails to train a global AI model efficiently. Accordingly, privacy-preserving distributed and collaborative machine learning, particularly Federated Learning (FL), is a desideratum. Already prevalent in the healthcare sector, questions remain regarding the effectiveness and efficacy of FL-based phishing detection within the context of multi-organization collaborations. To the best of our knowledge, the work herein is the first to investigate the use of FL in email anti-phishing. This paper builds upon a deep neural network model, particularly RNN and BERT for phishing email detection. It analyzes the FL-entangled learning performance under various settings, including balanced and asymmetrical data distribution. Our results corroborate comparable performance statistics of FL in phishing email detection to centralized learning for balanced datasets, and low organization counts. Moreover, we observe a variation in performance when increasing organizational counts. For a fixed total email dataset, the global RNN based model suffers by a 1.8% accuracy drop when increasing organizational counts from 2 to 10. In contrast, BERT accuracy rises by 0.6% when going from 2 to 5 organizations. However, if we allow increasing the overall email dataset with the introduction of new organizations in the FL framework, the organizational level performance is improved by achieving a faster convergence speed. Besides, FL suffers in its overall global model performance due to highly unstable outputs if the email dataset distribution is highly asymmetric.

Shuo Wang, Surya Nepal, Alsharif Abuadbba et al.

Machine learning models have demonstrated vulnerability to adversarial attacks, more specifically misclassification of adversarial examples. In this paper, we investigate an attack-agnostic defense against adversarial attacks on high-resolution images by detecting suspicious inputs. The intuition behind our approach is that the essential characteristics of a normal image are generally consistent with non-essential style transformations, e.g., slightly changing the facial expression of human portraits. In contrast, adversarial examples are generally sensitive to such transformations. In our approach to detect adversarial instances, we propose an in\underline{V}ertible \underline{A}utoencoder based on the \underline{S}tyleGAN2 generator via \underline{A}dversarial training (VASA) to inverse images to disentangled latent codes that reveal hierarchical styles. We then build a set of edited copies with non-essential style transformations by performing latent shifting and reconstruction, based on the correspondences between latent codes and style transformations. The classification-based consistency of these edited copies is used to distinguish adversarial instances.

Bushra Sabir, M. Ali Babar, Raj Gaire et al.

ML-based Phishing URL (MLPU) detectors serve as the first level of defence to protect users and organisations from being victims of phishing attacks. Lately, few studies have launched successful adversarial attacks against specific MLPU detectors raising questions about their practical reliability and usage. Nevertheless, the robustness of these systems has not been extensively investigated. Therefore, the security vulnerabilities of these systems, in general, remain primarily unknown which calls for testing the robustness of these systems. In this article, we have proposed a methodology to investigate the reliability and robustness of 50 representative state-of-the-art MLPU models. Firstly, we have proposed a cost-effective Adversarial URL generator URLBUG that created an Adversarial URL dataset. Subsequently, we reproduced 50 MLPU (traditional ML and Deep learning) systems and recorded their baseline performance. Lastly, we tested the considered MLPU systems on Adversarial Dataset and analyzed their robustness and reliability using box plots and heat maps. Our results showed that the generated adversarial URLs have valid syntax and can be registered at a median annual price of \$11.99. Out of 13\% of the already registered adversarial URLs, 63.94\% were used for malicious purposes. Moreover, the considered MLPU models Matthew Correlation Coefficient (MCC) dropped from a median 0.92 to 0.02 when tested against $Adv_\mathrm{data}$, indicating that the baseline MLPU models are unreliable in their current form. Further, our findings identified several security vulnerabilities of these systems and provided future directions for researchers to design dependable and secure MLPU systems.