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.
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
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.
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.
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.
