Michail Maniatakos

Esha Sarkar, Eduardo Chielle, Gamze Gursoy et al.

Machine Learning (ML) alleviates the challenges of high-dimensional data analysis and improves decision making in critical applications like healthcare. Effective cancer type from high-dimensional genetic mutation data can be useful for cancer diagnosis and treatment, if the distinguishable patterns between cancer types are identified. At the same time, analysis of high-dimensional data is computationally expensive and is often outsourced to cloud services. Privacy concerns in outsourced ML, especially in the field of genetics, motivate the use of encrypted computation, like Homomorphic Encryption (HE). But restrictive overheads of encrypted computation deter its usage. In this work, we explore the challenges of privacy preserving cancer detection using a real-world dataset consisting of more than 2 million genetic information for several cancer types. Since the data is inherently high-dimensional, we explore smaller ML models for cancer prediction to enable fast inference in the privacy preserving domain. We develop a solution for privacy preserving cancer inference which first leverages the domain knowledge on somatic mutations to efficiently encode genetic mutations and then uses statistical tests for feature selection. Our logistic regression model, built using our novel encoding scheme, achieves 0.98 micro-average area under curve with 13% higher test accuracy than similar studies. We exhaustively test our model's predictive capabilities by analyzing the genes used by the model. Furthermore, we propose a fast matrix multiplication algorithm that can efficiently handle high-dimensional data. Experimental results show that, even with 40,000 features, our proposed matrix multiplication algorithm can speed up concurrent inference of multiple individuals by approximately 10x and inference of a single individual by approximately 550x, in comparison to standard matrix multiplication.

Manaar Alam, Hithem Lamri, Michail Maniatakos

Federated Learning (FL) enables collaborative deep learning training across multiple participants without exposing sensitive personal data. However, the distributed nature of FL and the unvetted participants' data makes it vulnerable to backdoor attacks. In these attacks, adversaries inject malicious functionality into the centralized model during training, leading to intentional misclassifications for specific adversary-chosen inputs. While previous research has demonstrated successful injections of persistent backdoors in FL, the persistence also poses a challenge, as their existence in the centralized model can prompt the central aggregation server to take preventive measures to penalize the adversaries. Therefore, this paper proposes a methodology that enables adversaries to effectively remove backdoors from the centralized model upon achieving their objectives or upon suspicion of possible detection. The proposed approach extends the concept of machine unlearning and presents strategies to preserve the performance of the centralized model and simultaneously prevent over-unlearning of information unrelated to backdoor patterns, making the adversaries stealthy while removing backdoors. To the best of our knowledge, this is the first work that explores machine unlearning in FL to remove backdoors to the benefit of adversaries. Exhaustive evaluation considering image classification scenarios demonstrates the efficacy of the proposed method in efficient backdoor removal from the centralized model, injected by state-of-the-art attacks across multiple configurations.

Alejandro Guerra-Manzanares, L. Julian Lechuga Lopez, Michail Maniatakos et al.

Machine Learning (ML) has recently shown tremendous success in modeling various healthcare prediction tasks, ranging from disease diagnosis and prognosis to patient treatment. Due to the sensitive nature of medical data, privacy must be considered along the entire ML pipeline, from model training to inference. In this paper, we conduct a review of recent literature concerning Privacy-Preserving Machine Learning (PPML) for healthcare. We primarily focus on privacy-preserving training and inference-as-a-service, and perform a comprehensive review of existing trends, identify challenges, and discuss opportunities for future research directions. The aim of this review is to guide the development of private and efficient ML models in healthcare, with the prospects of translating research efforts into real-world settings.

Yue Wang, Wenqing Li, Esha Sarkar et al.

Backdoor attacks impose a new threat in Deep Neural Networks (DNNs), where a backdoor is inserted into the neural network by poisoning the training dataset, misclassifying inputs that contain the adversary trigger. The major challenge for defending against these attacks is that only the attacker knows the secret trigger and the target class. The problem is further exacerbated by the recent introduction of "Hidden Triggers", where the triggers are carefully fused into the input, bypassing detection by human inspection and causing backdoor identification through anomaly detection to fail. To defend against such imperceptible attacks, in this work we systematically analyze how representations, i.e., the set of neuron activations for a given DNN when using the training data as inputs, are affected by backdoor attacks. We propose PiDAn, an algorithm based on coherence optimization purifying the poisoned data. Our analysis shows that representations of poisoned data and authentic data in the target class are still embedded in different linear subspaces, which implies that they show different coherence with some latent spaces. Based on this observation, the proposed PiDAn algorithm learns a sample-wise weight vector to maximize the projected coherence of weighted samples, where we demonstrate that the learned weight vector has a natural "grouping effect" and is distinguishable between authentic data and poisoned data. This enables the systematic detection and mitigation of backdoor attacks. Based on our theoretical analysis and experimental results, we demonstrate the effectiveness of PiDAn in defending against backdoor attacks that use different settings of poisoned samples on GTSRB and ILSVRC2012 datasets. Our PiDAn algorithm can detect more than 90% infected classes and identify 95% poisoned samples.

Yue Wang, Wending Li, Michail Maniatakos et al.

Deep Reinforcement Learning (DRL) enhances the efficiency of Autonomous Vehicles (AV), but also makes them susceptible to backdoor attacks that can result in traffic congestion or collisions. Backdoor functionality is typically incorporated by contaminating training datasets with covert malicious data to maintain high precision on genuine inputs while inducing the desired (malicious) outputs for specific inputs chosen by adversaries. Current defenses against backdoors mainly focus on image classification using image-based features, which cannot be readily transferred to the regression task of DRL-based AV controllers since the inputs are continuous sensor data, i.e., the combinations of velocity and distance of AV and its surrounding vehicles. Our proposed method adds well-designed noise to the input to neutralize backdoors. The approach involves learning an optimal smoothing (noise) distribution to preserve the normal functionality of genuine inputs while neutralizing backdoors. By doing so, the resulting model is expected to be more resilient against backdoor attacks while maintaining high accuracy on genuine inputs. The effectiveness of the proposed method is verified on a simulated traffic system based on a microscopic traffic simulator, where experimental results showcase that the smoothed traffic controller can neutralize all trigger samples and maintain the performance of relieving traffic congestion

Bilel Sefsaf, Abderraouf Dandani, Abdessamed Seddiki et al.

Fully Homomorphic Encryption (FHE) enables computations directly on encrypted data, but its high computational cost remains a significant barrier. Writing efficient FHE code is a complex task requiring cryptographic expertise, and finding the optimal sequence of program transformations is often intractable. In this paper, we propose CHEHAB RL, a novel framework that leverages deep reinforcement learning (RL) to automate FHE code optimization. Instead of relying on predefined heuristics or combinatorial search, our method trains an RL agent to learn an effective policy for applying a sequence of rewriting rules to automatically vectorize scalar FHE code while reducing instruction latency and noise growth. The proposed approach supports the optimization of both structured and unstructured code. To train the agent, we synthesize a diverse dataset of computations using a large language model (LLM). We integrate our proposed approach into the CHEHAB FHE compiler and evaluate it on a suite of benchmarks, comparing its performance against Coyote, a state-of-the-art vectorizing FHE compiler. The results show that our approach generates code that is $5.3\times$ faster in execution, accumulates $2.54\times$ less noise, while the compilation process itself is $27.9\times$ faster than Coyote (geometric means).

Hithem Lamri, Michail Maniatakos

Machine unlearning (MU) seeks to remove the influence of specified data from a trained model in response to privacy requests or data poisoning. While certified unlearning has been analyzed in centralized and server-orchestrated federated settings (via guarantees analogous to differential privacy, DP), the decentralized setting -- where peers communicate without a coordinator remains underexplored. We study certified unlearning in decentralized networks with fixed topologies and propose RR-DU, a random-walk procedure that performs one projected gradient ascent step on the forget set at the unlearning client and a geometrically distributed number of projected descent steps on the retained data elsewhere, combined with subsampled Gaussian noise and projection onto a trust region around the original model. We provide (i) convergence guarantees in the convex case and stationarity guarantees in the nonconvex case, (ii) $(\varepsilon,δ)$ network-unlearning certificates on client views via subsampled Gaussian Rényi DP (RDP) with segment-level subsampling, and (iii) deletion-capacity bounds that scale with the forget-to-local data ratio and quantify the effect of decentralization (network mixing and randomized subsampling) on the privacy-utility trade-off. Empirically, on image benchmarks (MNIST, CIFAR-10), RR-DU matches a given $(\varepsilon,δ)$ while achieving higher test accuracy than decentralized DP baselines and reducing forget accuracy to random guessing ($\approx 10\%$).

Manaar Alam, Hithem Lamri, Michail Maniatakos

Backdoor attacks embed hidden functionalities in deep neural networks (DNN), triggering malicious behavior with specific inputs. Advanced defenses monitor anomalous DNN inferences to detect such attacks. However, concealed backdoors evade detection by maintaining a low pre-deployment attack success rate (ASR) and restoring high ASR post-deployment via machine unlearning. Existing concealed backdoors are often constrained by requiring white-box or black-box access or auxiliary data, limiting their practicality when such access or data is unavailable. This paper introduces ReVeil, a concealed backdoor attack targeting the data collection phase of the DNN training pipeline, requiring no model access or auxiliary data. ReVeil maintains low pre-deployment ASR across four datasets and four trigger patterns, successfully evades three popular backdoor detection methods, and restores high ASR post-deployment through machine unlearning.

Alejandro Guerra-Manzanares, Omar El-Herraoui, Michail Maniatakos et al.

One of the key challenges of collaborative machine learning, without data sharing, is multimodal data heterogeneity in real-world settings. While Federated Learning (FL) enables model training across multiple clients, existing frameworks, such as horizontal and vertical FL, are only effective in `ideal' settings that meet specific assumptions. Hence, they struggle to address scenarios where neither all modalities nor all samples are represented across the participating clients. To address this gap, we propose BlendFL, a novel FL framework that seamlessly blends the principles of horizontal and vertical FL in a synchronized and non-restrictive fashion despite the asymmetry across clients. Specifically, any client within BlendFL can benefit from either of the approaches, or both simultaneously, according to its available dataset. In addition, BlendFL features a decentralized inference mechanism, empowering clients to run collaboratively trained local models using available local data, thereby reducing latency and reliance on central servers for inference. We also introduce BlendAvg, an adaptive global model aggregation strategy that prioritizes collaborative model updates based on each client's performance. We trained and evaluated BlendFL and other state-of-the-art baselines on three classification tasks using a large-scale real-world multimodal medical dataset and a popular multimodal benchmark. Our results highlight BlendFL's superior performance for both multimodal and unimodal classification. Ablation studies demonstrate BlendFL's faster convergence compared to traditional approaches, accelerating collaborative learning. Overall, in our study we highlight the potential of BlendFL for handling multimodal data heterogeneity for collaborative learning in real-world settings where data privacy is crucial, such as in healthcare and finance.

Hithem Lamri, Manaar Alam, Haiyan Jiang et al.

Federated Unlearning (FU) enables clients to remove the influence of specific data from a collaboratively trained shared global model, addressing regulatory requirements such as GDPR and CCPA. However, this unlearning process introduces a new privacy risk: A malicious server may exploit unlearning updates to reconstruct the data requested for removal, a form of Data Reconstruction Attack (DRA). While DRAs for machine unlearning have been studied extensively in centralized Machine Learning-as-a-Service (MLaaS) settings, their applicability to FU remains unclear due to the decentralized, client-driven nature of FU. This work presents DRAUN, the first attack framework to reconstruct unlearned data in FU systems. DRAUN targets optimization-based unlearning methods, which are widely adopted for their efficiency. We theoretically demonstrate why existing DRAs targeting machine unlearning in MLaaS fail in FU and show how DRAUN overcomes these limitations. We validate our approach through extensive experiments on four datasets and four model architectures, evaluating its performance against five popular unlearning methods, effectively demonstrating that state-of-the-art FU methods remain vulnerable to DRAs.

Christoforos Vasilatos, Dunia J. Mahboobeh, Hithem Lamri et al.

Industrial Control Systems (ICS) are extensively used in critical infrastructures ensuring efficient, reliable, and continuous operations. However, their increasing connectivity and addition of advanced features make them vulnerable to cyber threats, potentially leading to severe disruptions in essential services. In this context, honeypots play a vital role by acting as decoy targets within ICS networks, or on the Internet, helping to detect, log, analyze, and develop mitigations for ICS-specific cyber threats. Deploying ICS honeypots, however, is challenging due to the necessity of accurately replicating industrial protocols and device characteristics, a crucial requirement for effectively mimicking the unique operational behavior of different industrial systems. Moreover, this challenge is compounded by the significant manual effort required in also mimicking the control logic the PLC would execute, in order to capture attacker traffic aiming to disrupt critical infrastructure operations. In this paper, we propose LLMPot, a novel approach for designing honeypots in ICS networks harnessing the potency of Large Language Models (LLMs). LLMPot aims to automate and optimize the creation of realistic honeypots with vendor-agnostic configurations, and for any control logic, aiming to eliminate the manual effort and specialized knowledge traditionally required in this domain. We conducted extensive experiments focusing on a wide array of parameters, demonstrating that our LLM-based approach can effectively create honeypot devices implementing different industrial protocols and diverse control logic.

Christoforos Vasilatos, Manaar Alam, Talal Rahwan et al.

As the use of Large Language Models (LLMs) in text generation tasks proliferates, concerns arise over their potential to compromise academic integrity. The education sector currently tussles with distinguishing student-authored homework assignments from AI-generated ones. This paper addresses the challenge by introducing HowkGPT, designed to identify homework assignments generated by AI. HowkGPT is built upon a dataset of academic assignments and accompanying metadata [17] and employs a pretrained LLM to compute perplexity scores for student-authored and ChatGPT-generated responses. These scores then assist in establishing a threshold for discerning the origin of a submitted assignment. Given the specificity and contextual nature of academic work, HowkGPT further refines its analysis by defining category-specific thresholds derived from the metadata, enhancing the precision of the detection. This study emphasizes the critical need for effective strategies to uphold academic integrity amidst the growing influence of LLMs and provides an approach to ensuring fair and accurate grading in educational institutions.

Hazem Ibrahim, Fengyuan Liu, Rohail Asim et al.

The emergence of large language models has led to the development of powerful tools such as ChatGPT that can produce text indistinguishable from human-generated work. With the increasing accessibility of such technology, students across the globe may utilize it to help with their school work -- a possibility that has sparked discussions on the integrity of student evaluations in the age of artificial intelligence (AI). To date, it is unclear how such tools perform compared to students on university-level courses. Further, students' perspectives regarding the use of such tools, and educators' perspectives on treating their use as plagiarism, remain unknown. Here, we compare the performance of ChatGPT against students on 32 university-level courses. We also assess the degree to which its use can be detected by two classifiers designed specifically for this purpose. Additionally, we conduct a survey across five countries, as well as a more in-depth survey at the authors' institution, to discern students' and educators' perceptions of ChatGPT's use. We find that ChatGPT's performance is comparable, if not superior, to that of students in many courses. Moreover, current AI-text classifiers cannot reliably detect ChatGPT's use in school work, due to their propensity to classify human-written answers as AI-generated, as well as the ease with which AI-generated text can be edited to evade detection. Finally, we find an emerging consensus among students to use the tool, and among educators to treat this as plagiarism. Our findings offer insights that could guide policy discussions addressing the integration of AI into educational frameworks.

Constantine Doumanidis, Prashant Hari Narayan Rajput, Michail Maniatakos

Industrial Control Systems (ICS) have played a catalytic role in enabling the 4th Industrial Revolution. ICS devices like Programmable Logic Controllers (PLCs), automate, monitor, and control critical processes in industrial, energy, and commercial environments. The convergence of traditional Operational Technology (OT) with Information Technology (IT) has opened a new and unique threat landscape. This has inspired defense research that focuses heavily on Machine Learning (ML) based anomaly detection methods that run on external IT hardware, which means an increase in costs and the further expansion of the threat landscape. To remove this requirement, we introduce the ICS machine learning inference framework (ICSML) which enables executing ML model inference natively on the PLC. ICSML is implemented in IEC 61131-3 code and provides several optimizations to bypass the limitations imposed by the domain-specific languages. Therefore, it works on every PLC without the need for vendor support. ICSML provides a complete set of components for creating full ML models similarly to established ML frameworks. We run a series of benchmarks studying memory and performance, and compare our solution to the TFLite inference framework. At the same time, we develop domain-specific model optimizations to improve the efficiency of ICSML. To demonstrate the abilities of ICSML, we evaluate a case study of a real defense for process-aware attacks targeting a desalination plant.

Esha Sarkar, Michail Maniatakos

Machine Learning (ML) has achieved unprecedented performance in several applications including image, speech, text, and data analysis. Use of ML to understand underlying patterns in gene mutations (genomics) has far-reaching results, not only in overcoming diagnostic pitfalls, but also in designing treatments for life-threatening diseases like cancer. Success and sustainability of ML algorithms depends on the quality and diversity of data collected and used for training. Under-representation of groups (ethnic groups, gender groups, etc.) in such a dataset can lead to inaccurate predictions for certain groups, which can further exacerbate systemic discrimination issues. In this work, we propose TRAPDOOR, a methodology for identification of biased datasets by repurposing a technique that has been mostly proposed for nefarious purposes: Neural network backdoors. We consider a typical collaborative learning setting of the genomics supply chain, where data may come from hospitals, collaborative projects, or research institutes to a central cloud without awareness of bias against a sensitive group. In this context, we develop a methodology to leak potential bias information of the collective data without hampering the genuine performance using ML backdooring catered for genomic applications. Using a real-world cancer dataset, we analyze the dataset with the bias that already existed towards white individuals and also introduced biases in datasets artificially, and our experimental result show that TRAPDOOR can detect the presence of dataset bias with 100% accuracy, and furthermore can also extract the extent of bias by recovering the percentage with a small error.

Munachiso Nwadike, Takumi Miyawaki, Esha Sarkar et al.

Deep neural networks have been shown to be vulnerable to backdoor attacks, which could be easily introduced to the training set prior to model training. Recent work has focused on investigating backdoor attacks on natural images or toy datasets. Consequently, the exact impact of backdoors is not yet fully understood in complex real-world applications, such as in medical imaging where misdiagnosis can be very costly. In this paper, we explore the impact of backdoor attacks on a multi-label disease classification task using chest radiography, with the assumption that the attacker can manipulate the training dataset to execute the attack. Extensive evaluation of a state-of-the-art architecture demonstrates that by introducing images with few-pixel perturbations into the training set, an attacker can execute the backdoor successfully without having to be involved with the training procedure. A simple 3$\times$3 pixel trigger can achieve up to 1.00 Area Under the Receiver Operating Characteristic (AUROC) curve on the set of infected images. In the set of clean images, the backdoored neural network could still achieve up to 0.85 AUROC, highlighting the stealthiness of the attack. As the use of deep learning based diagnostic systems proliferates in clinical practice, we also show how explainability is indispensable in this context, as it can identify spatially localized backdoors in inference time.

Esha Sarkar, Hadjer Benkraouda, Michail Maniatakos

Recent advances in Machine Learning (ML) have opened up new avenues for its extensive use in real-world applications. Facial recognition, specifically, is used from simple friend suggestions in social-media platforms to critical security applications for biometric validation in automated immigration at airports. Considering these scenarios, security vulnerabilities to such ML algorithms pose serious threats with severe outcomes. Recent work demonstrated that Deep Neural Networks (DNNs), typically used in facial recognition systems, are susceptible to backdoor attacks; in other words,the DNNs turn malicious in the presence of a unique trigger. Adhering to common characteristics for being unnoticeable, an ideal trigger is small, localized, and typically not a part of the main im-age. Therefore, detection mechanisms have focused on detecting these distinct trigger-based outliers statistically or through their reconstruction. In this work, we demonstrate that specific changes to facial characteristics may also be used to trigger malicious behavior in an ML model. The changes in the facial attributes maybe embedded artificially using social-media filters or introduced naturally using movements in facial muscles. By construction, our triggers are large, adaptive to the input, and spread over the entire image. We evaluate the success of the attack and validate that it does not interfere with the performance criteria of the model. We also substantiate the undetectability of our triggers by exhaustively testing them with state-of-the-art defenses.

Yue Wang, Esha Sarkar, Wenqing Li et al.

Recent work has shown that the introduction of autonomous vehicles (AVs) in traffic could help reduce traffic jams. Deep reinforcement learning methods demonstrate good performance in complex control problems, including autonomous vehicle control, and have been used in state-of-the-art AV controllers. However, deep neural networks (DNNs) render automated driving vulnerable to machine learning-based attacks. In this work, we explore the backdooring/trojanning of DRL-based AV controllers. We develop a trigger design methodology that is based on well-established principles of traffic physics. The malicious actions include vehicle deceleration and acceleration to cause stop-and-go traffic waves to emerge (congestion attacks) or AV acceleration resulting in the AV crashing into the vehicle in front (insurance attack). We test our attack on single-lane and two-lane circuits. Our experimental results show that the backdoored model does not compromise normal operation performance, with the maximum decrease in cumulative rewards being 1%. Still, it can be maliciously activated to cause a crash or congestion when the corresponding triggers appear.

Anastasis Keliris, Michail Maniatakos

The security of Industrial Control Systems (ICS) has been attracting increased attention over the past years, following the discovery of real threats targeting industrial environments. Despite this attention, automation of the reverse engineering process of ICS binaries for programmable logic controllers remains an open problem, mainly due to the use of proprietary compilers by ICS vendors. Such automation could be a double-edged sword; on the one hand it could accelerate digital forensic investigations and incident response actions, while on the other hand it could enable dynamic generation of malicious ICS payloads. In this work, we propose a structured methodology that automates the reverse engineering process for ICS binaries taking into account their unique domain-specific characteristics. We apply this methodology to develop the modular Industrial Control Systems Reverse Engineering Framework (ICSREF), and instantiate ICSREF modules for reversing binaries compiled with CODESYS, a widely used software stack and compiler for PLCs. To evaluate our framework we create a database of samples by collecting real PLC binaries from public code repositories, as well as developing binaries in-house. Our results demonstrate that ICSREF can successfully handle diverse PLC binaries from varied industry sectors, irrespective of the programming language used. Furthermore, we deploy ICSREF on a commercial smartphone which orchestrates and launches a completely automated process-aware attack against a chemical process testbed. This example of dynamic payload generation showcases how ICSREF can enable sophisticated attacks without any prior knowledge.