Advanced Persistent Threat (APT) is a type of cyberattack where intruders gain unauthorized access and remain undetected for an elongated period of time. The main purpose of an APT is to steal sensitive information, spy on operations, or disrupt systems. To overcome these challenges, in this paper, we have proposed a novel modular and scalable architecture SANVector (SBERT-APTNet Vector) framework that uses fine-tuned SBERT embeddings, contrastive learning, and multi-view features to automatically identify the threat actors from unstructured reports about cyber threats. For that, we have focused on attributing APT activities to known threat actor groups such as APT28, APT29, Lazarus Group, Turla, and others extracted from various log report sources such as FireEye, Crowdstrike, Unit42, Mandiant, and Kaspersky. In all we analyze a total of 11 APT groups. To begin, our model incorporates a strong preprocessing pipeline that normalizes attacker aliases, extracts deep semantic features, and masks indicators of compromises using a trained SBERT model. Furthermore, these features are classified using the traditional model with K-Nearest Neighbor (KNN) to achieve high performances of 93.17
Deepfakes have become commonplace due to the rise of GANs and LLM based generative models. With this, there has also been a rise in research towards effective deepfake-detection. In this work, we propose a transformer based model for deepfake-detection, operating on frequency-domain image features. Alongside, we also test the effectiveness of our solution with a smaller CNN-based model, for the sake of comparing model complexity vis-a-vis feature set effectiveness. In this paper, we focus on detection of Diffusion Model (DM) generated deepfake images. We train our models on an extensive dataset of diffusion model–generated images. We also demonstrate their generalization capabilities and report that while smaller models work well for in-domain testing, we do need larger, deeper neural networks for improved generalization. Our experiments on the recently proposed DeepFakeEval dataset corroborate our findings.
We present three new algorithms for efficient in-place estimation of average fidelity of a d-dimensional quantum logic gate, without using ancilla qubits. The main advantage of our algorithms is the much smaller usage of truly random bits compared to what was known so far. Reducing the requirement of classical seed randomness increases the reliability of estimation, as high-quality random bits are an expensive computational resource and commonly used sources of random bits are either pseudorandom or have a non-trivial bias. We use cutting edge tools from classical and quantum derandomisation theory, viz. k-wise independent generators and k-fold quantum tensor product expanders in order to obtain our new randomness efficient algorithms. Our algorithms are suitable for NISQ and early fault-tolerant quantum computers where one wants to reduce the number of control pulses that can potentially be used, in order to minimise the complexity of the control circuitry and also due to coherence time limitations.
Text summarization is an important problem in natural language processing. The advent of large-scale transformer-based models, such as BART, PEGASUS, and T5, has revolutionized abstractive summarization. However, all these models are centralized and require the entire text to be summarized at a single location, where the model operates and generates the summary. Centralized training of NLP models raises privacy concerns in sensitive domains such as healthcare, law, and finance. Another challenge concerning these models is their unfair representation of users with small corpus data. To mitigate this, we propose a privacy-preserving framework for abstractive summarization using federated learning, eliminating the need to transmit raw data to a centralized server. Our approach employs a pre-trained BART model, and to ensure privacy against inference attacks by the aggregating server, differential privacy is utilized. Our approach also incorporates a fairness-aware aggregation strategy, Q-fair, to ensure balanced client contributions. Experiments on the CNN/DailyMail dataset show that our FL-based method achieves competitive ROUGE scores while enhancing privacy and fairness, demonstrating its practicality for sensitive real-world applications.
Privacy-Preserving classification of Machine Learning (PPML) models has gained traction, owing to strict data privacy regulations. One such work, CryptoEyes (Wenbo et al., INFOCOM 2021) enables users to encrypt images using AES in ECB mode and store it at the server for processing. The authors claimed that their protocol have two advantages: (i) service provider could extract enough information in the form of (encrypted) contour, which gives it the ability to perform machine learning classification on encrypted images, and (ii) use of secret permutation known only to user and server, this prevents man-in-the-middle adversary from reconstructing meaningful image content or accurately map encrypted data to its original class distribution even after having access to a train dataset, thereby achieving privacy. However, in this paper we demonstrate an adversarial attack, which surpass claimed classification accuracy for an adversary against CryptoEyes. Our attack approach uses new data transformation techniques for better feature identification and then train an ML model on image pixels encrypted and permuted. We identify the key vulnerabilities enabling this data leakage and show how an adversary can exploit these weaknesses. The training of our learning model is key agnostic. The results show that with our attack the adversary is able to train a model on MNIST dataset with at least 74.2% classification accuracy on encrypted and permuted image pixels.
Neural network models have been widely used to make critical decisions across diverse applications. However, these models are susceptible to inference attacks that can expose sensitive information such as gender, race, and other personal attributes from the private data used during training. We focus on a particular variant of this attack, known as sensitive value inference, where the adversary aims to reliably pinpoint records within a candidate pool that possess a specific value for the sensitive attribute. We exploit neuron activation values to frame this attack, to infer sensitive attributes. The attack is based on the observation that some neurons are strongly correlated with specific sensitive attribute values in the input. After identifying the most important neurons, the attacker selects the top-k and uses their activation values to train an attack model that can predict the sensitive attribute. We construct the attack based on two distinct threat scenarios: (a) where the model creator incorporates sensitive attributes in both the training data and model inputs, and (b) where these sensitive attributes are intentionally excluded from both the training data and input to censor them. We evaluated our attack on the COMPAS, CENSUS, and Texas-100x, and UCI credit card datasets. It is observed that the proposed attack achieves an inference precision of 72% when the sensitive attribute is omitted from the training data, and this precision increases to 93% when the sensitive attribute is included during training.
A smart grid is a modern power network designed to monitor, manage, and optimise the flow of electricity. A cyber attack on such a critical system can disrupt services and pose significant risks to public safety and national infrastructure. Cyber attackers targeting critical infrastructures like smart grids often follow predictable patterns that can be modelled and analysed. Attack graphs reveal these potential multi-stage attack patterns, highlighting attacker reachability. However, due to the sheer number of paths and the large size of the graph, their analysis becomes computationally challenging. In this paper, a modular analysis of attack graphs is proposed to address this issue by decomposing the graphs into manageable clusters or communities. By representing each community as a single node, the complex graph is transformed into a high-level abstraction that is easier to interpret. Segregating large attack graphs into clusters enables efficient, time-bound analysis and ultimately aids in defense strategy formulation.
Quantum computing poses a major threat to classical cryptographic primitives (RSA, ECC, etc.), demanding an urgent migration to quantum-safe solutions, in any sensitive IT sectors such as banking enterprises. Merely replacing classical cryptographic primitives does not fully address this migration. This migration requires upgrades to network devices (e.g., routers, firewalls, servers) to ensure compatibility with future quantum-safe security protocols. Due to the scale of enterprise networks, a phased upgrade is necessary, as simultaneous migration of all devices is infeasible. However, a quantitative method for identifying network device vulnerabilities in large-scale enterprise networks is currently lacking. We propose DoPQM: Devices Oriented Post-Quantum Migration, a framework for identifying network device vulnerabilities to quantum threats within enterprises network topologies. The primary objective of this work is to assist network administrators in identifying which network devices are more vulnerable to future quantum threats. DoPQM formalizes the Harvest Now, Decrypt Later (HNDL) adversarial model using a risk scoring mechanism that estimates HNDL attack likelihoods across network nodes. By leveraging graph theory, we identify central nodes based on the volume of data they handle. Each node is then assigned a criticality score that reflects its priority for post-quantum migration. Evaluation on a representative baking network reveals the Demilitarized Zone (DMZ) as the most HNDL vulnerable component.
The automation of sensitive information classification using AI is critical for modern security, yet a singular focus on accuracy metrics dangerously obscures the risks posed by residual errors. False Negatives can lead to catastrophic data breaches, creating an unmanaged attack surface. This paper bridges the gap between formal security theory and applied machine learning by proposing a framework to operationalize principles of conditional secrecy. We demonstrate how the confidence scores from Transformer models can be used as a practical mechanism to enforce a principled secrecy policy, segmenting data into dynamic security levels. This approach allows organizations to manage the trade-off between automation coverage and risk exposure by treating low-confidence classifications as a controlled transfer to a trusted group of human reviewers. We validate this framework on the expert-annotated Monsanto Papers corpus, showing that Transformer-based classification, unlike traditional methods, provides the necessary discriminative power to make this principled approach viable. Our work offers a new, risk-aware methodology for the secure deployment of AI in high-stakes information governance.
As open-source solutions are increasingly deployed in 5G Radio Access Networks (RANs), the importance of reliable security analysis for these systems is growing - especially for popular frameworks like srsRAN. Due to the complex and performance-critical nature of RAN software, latent vulnerabilities can have serious consequences for user privacy and network integrity. In this work, we present a detailed static analysis of the srsRAN codebase using CodeQL, a semantic code query language designed to scale vulnerability detection across large codebases. By formulating and executing custom CodeQL queries, our analysis detects a variety of systemic problems in srsRAN, such as memory management errors (e.g., memory leaks, use-after-free, null pointer dereferences), dangerous system calls, inadequate access control enforcement, unchecked return values, buffer overflows, and unvalidated network inputs. Other issues uncovered include hardcoded secrets, use of cryptographically weak pseudorandom number generation, and unsafe cryptographic comparisons - flaws that could potentially lead to privilege escalation, remote code execution, or denial-of-service attacks. The results highlight the importance of static code analysis tools like CodeQL for proactively hardening telecommunications infrastructure, and encourage the adoption of code analysis techniques in the secure software development lifecycle of open-source 5G RAN implementations.
The inherent complexity of cloud-native environments like Kubernetes is due to its dynamic and distributed landscape. According to the Kubernetes Security Report 2024, over 67
Collaborative edge data analytics is gaining prominence as vast amounts of data are increasingly generated at the network edge by smartphones, sensors, and smart cameras. In this work, we focus on the challenge of privacy-preserving video analytics on edge devices. Video data presents unique difficulties compared to other data types, primarily due to its large size and the nature of video analysis tasks. We evaluate and compare two leading techniques for enabling secure collaborative video analytics in distributed edge environments: Secure Multiparty Computation (MPC) and Trusted Execution Environments (TEE). To assess their effectiveness, we implement five real-world case studies, including object re-identification, scene similarity detection, vehicle counting, and machine learning–based video fusion tasks. Additionally, we benchmark fundamental image processing operations under various MPC configurations to identify the most efficient MPC protocols for video workloads. Our results show that while TEE offer significant performance benefits, especially for machine learning intensive tasks, MPC remains a practical alternative in scenarios without trusted hardware, particularly when using optimized secret sharing based 3-party protocols. We provide a comprehensive analysis of performance, security, and implementation complexity for both approaches.
We present a new pattern generation method for the Logical Analysis of Data (LAD) that is based on genetic algorithms. The original LAD method is often too slow and struggles to find patterns with many features. Our approach makes it faster and allows the model to find more complex patterns, even if it may not always find the perfect ones. We tested this method in credit card fraud detection, which is the process of identifying hidden or fraudulent transactions among many real ones. This is important because missing fraud or falsely labeling real transactions can cause serious problems. Our results show that the proposed method demonstrates superior performance across various dataset configurations. On a 3:1 sampled dataset, it achieves 97.46
Machine Learning-as-a-Service (MLaaS) enables deep learning (DL) model owners to outsource inference tasks to a public cloud platform. A model owner trains a DL model in-house and uploads the trained model to a MLaaS. For the uploaded model, the MLaaS platform exposes an API to query the uploaded model with inputs and obtain predictions. However, uploading the trained model to public cloud platforms exposes the model owner to security and privacy risks, as the model is available in plaintext to the cloud provider during inference. In this work, we present techniques to secure DL models with trusted execution environments and propose a secure outsourcing scheme to offload portions of the DL model computations during inference to faster untrusted processors. We implement the presented techniques in MazeNet, a framework to transform pretrained models into MazeNet models and deploy them on a public cloud platform to provide inference services. We evaluate MazeNet on popular convolutional neural networks, and the results demonstrate that MazeNet improves the performance of DNN models as compared to a secure baseline model, where the model runs within a trusted environment. MazeNet increases the throughput of the inference task up to 30x and decreases the latency up to 5x for the benchmark models in our experimental evaluation.
The existing MQTT honeypots often lack scalability, semantic fidelity, and data-flow awareness, thereby limiting interaction realism and detection depth. To address these limitations, we propose an Extended Finite State Machine (EFSM) based MQTT honeypot that enforces guarded control and data-flow transitions while integrating adaptive deception to ensure protocol compliant realism. Unlike conventional honeypots, the EFSM based design captures field level dependencies through variable updates and guard conditions, enabling stateful, context aware emulation. In an 80 day Internet deployment, the proposed system achieved 23 1.5× higher topic diversity, and 40
Federated learning (FL) stands out as a promising paradigm for collaborative training of machine learning models where a server supervises the learning process while keeping sensitive data on the user devices. Here, training is decentralized and conducted on edge devices beyond the control of a server. This increases the potential for malicious clients to tamper with the learning process and compromise the global model, resulting in a significant security risk. The majority of existing solutions are designed for scenarios where data exhibits independent and identically distributed (IID) characteristics across devices. A notable performance degradation is observed when the data distribution deviates from the independent and identically distributed (non-IID) scenario. In this paper, we first evaluate the performance of existing Byzantine robust aggregation schemes in non-IID settings within an adversarial scenario. Then, we introduce a novel attack-resilient aggregation scheme named FedResil with the objective of enhancing performance in the same adversarial environment. It leverages non-private data, which is collectively agreed upon by the participating clients before the training process begins, to delineate clusters of clients. Subsequently, the server applies existing Byzantine robust aggregation rules to each cluster independently, generating model updates within each cluster. The model update from each cluster is then aggregated to construct the final global model. Through extensive experimentation, we demonstrate that FedResil in malicious settings achieves performance similar to that in scenarios where there is no malicious client.
Outsourced Ciphertext-Policy Attribute-Based Encryption (CP-ABE) has emerged as a promising solution that enables fine-grained access control over encrypted data, making it well-suited for cloud-based data storage and sharing systems, especially when resource-constrained devices delegate computationally intensive tasks to cloud servers. To address the computational overhead of encryption, several outsourced CP-ABE schemes have been proposed, allowing heavy encryption computations to be offloaded to cloud servers. However, outsourcing encryption to potentially untrusted cloud servers introduces new security challenges, particularly regarding data confidentiality. Recently, Zhang et al. and Miao et al. independently proposed two outsourced CP-ABE schemes. The Zhang et al. scheme is a proxy re-encryption scheme, where the cloud server converts an identity-based encryption ciphertext into a CP-ABE ciphertext to establish a fine-grained data sharing mechanism in the Industrial Internet of Things environments. The scheme proposed by Miao et al. is an outsourced CP-ABE with verifiable encryption that establishes a fine-grained data sharing mechanism in the cloud-assisted mobile electronic health system, and later they extend it for verifiable decryption. The authors in both the schemes claimed that their proposed schemes preserve data confidentiality against the semi-trusted cloud server and unauthorized users. However, after a comprehensive security analysis of these two prominent outsourced CP-ABE schemes, we identify that the two schemes fail to provide data confidentiality. After carefully revisiting Zhang et al. and Miao et al. schemes, in this paper, we demonstrate that their schemes fail to provide data confidentiality by proposing message recovery attacks. Our attacks show that the semi-trusted cloud server can successfully extract the plaintext encoded in a ciphertext.
Machine learning-based access control (MLBAC) shows promise in effectively determining access in complex scenarios where a trained ML model makes decisions. When access policies change, the underlying ML model must be updated to accommodate these changes. This process requires a portion of past training data, known as Replay Data, along with the new changes to retain existing knowledge and avoid challenges like catastrophic forgetting. Traditionally, Replay Data is selected randomly, constituting a large portion (approximately 25
To address the growing demand for robust Network Intrusion Detection Systems (NIDS), we propose a stage-wise curriculum learning framework combined with image-based transformation techniques and Explainable AI (XAI). This approach leverages XAI for transparency and comprehensibility, focusing on adaptive learning and scalable detection. By training the model on increasingly complex attack scenarios, the proposed architecture facilitates efficient and precise optimization. The explanatory tool SHAP (SHapley Additive exPlanations) is integrated provide detailed insights into model predictions. The ensemble stacking and model-to-model interoperability further enhances detection efficiency and reliability. Experiments validate the effectiveness of the proposed method, achieving the following accuracies: 97
It is well known that Cyber-Physical Systems (CPSs) are vulnerable to cyberattacks, and detection is usually achieved using machine learning-based Intrusion Detection Systems (IDSs). CPSs frequently have highly complex control logic, which, when maliciously manipulated, may lead to accidents. As control logic is not publicly available, here we propose a framework that can extract approximate control logic from the captured operational data flowing in the communication channel of a CPS. This approximate control logic is used to design stealthy adversarial samples with a very low footprint in terms of the number of sensor readings perturbed and the amount of sensor readings perturbed during sample generation. The proposed control logic induction framework involves four steps: cut-point generation, cut-point selection, boolean control rule extraction, and combining extracted control rules into control logic. Later, a control-logic-based stealthy adversarial sample generation technique was designed where the adversary possesses no knowledge regarding the targeted IDSs or their training data. Existing IDSs fail to recognize the stealthy adversarial samples in more than 98% of the cases. Even retraining the IDSs with adversarial samples generated using methods like GAN, FGSM, etc., fails to improve their resilience against stealthy attacks. On the contrary, retraining with stealthy samples improves the performance of IDSs not only against stealthy attacks but also against other types of adversarial attacks.
