Abhishek Mishra

Abhishek K. Mishra, Antoine Boutet, Lucas Magnana

Large Language Models (LLMs) are increasingly deployed across multilingual applications that handle sensitive data, yet their scale and linguistic variability introduce major privacy risks. Mostly evaluated for English, this paper investigates how language structure affects privacy leakage in LLMs trained on English, Spanish, French, and Italian medical corpora. We quantify six linguistic indicators and evaluate three attack vectors: extraction, counterfactual memorization, and membership inference. Results show that privacy vulnerability scales with linguistic redundancy and tokenization granularity: Italian exhibits the strongest leakage, while English shows higher membership separability. In contrast, French and Spanish display greater resilience due to higher morphological complexity. Overall, our findings provide the first quantitative evidence that language matters in privacy leakage, underscoring the need for language-aware privacy-preserving mechanisms in LLM deployments.

Abhishek Kumar Mishra, Aline Carneiro Viana, Nadjib Achir

Bluetooth has become critical as many IoT devices are arriving in the market. Most of the current literature focusing on Bluetooth simulation concentrates on the network protocols' performances and completely neglects the privacy protection recommendations introduced in the BLE standard. Indeed, privacy protection is one of the main issues handled in the Bluetooth standard. For instance, the current standard forces devices to change the identifier they embed within the public and private packets, known as MAC address randomization. Although randomizing MAC addresses is intended to preserve device privacy, recent literature shows many challenges that are still present. One of them is the correlation between the public packets and the emitters. Unfortunately, existing evaluation tools such as NS-3 are not designed to reproduce this Bluetooth standard's essential functionality. This makes it impossible to test solutions for different device-fingerprinting strategies as there is a lack of ground truth for large-scale scenarios with the majority of current BLE devices implementing MAC address randomization. In this paper, we first introduce a solution of standard-compliant MAC address randomization in the NS-3 framework, capable of emulating any real BLE device in the simulation and generating real-world Bluetooth traces. In addition, since the simulation run-time for trace-collection grows exponentially with the number of devices, we introduce an optimization to linearize public-packet sniffing. This made the large-scale trace-collection practically feasible. Then, we use the generated traces and associated ground truth to do a case study on the evaluation of a generic MAC address association available in the literature. Our case study reveals that close to 90 percent of randomized addresses could be correctly linked even in highly dense and mobile scenarios.

Sai Hemanth Kasaraneni, Abhishek Mishra

Image Completion refers to the task of filling in the missing regions of an image and Image Extrapolation refers to the task of extending an image at its boundaries while keeping it coherent. Many recent works based on GAN have shown progress in addressing these problem statements but lack adaptability for these two cases, i.e. the neural network trained for the completion of interior masked images does not generalize well for extrapolating over the boundaries and vice-versa. In this paper, we present a technique to train both completion and extrapolation networks concurrently while benefiting each other. We demonstrate our method's efficiency in completing large missing regions and we show the comparisons with the contemporary state of the art baseline.

Pratyusa Mukherjee, Krishnendu Rarhi, Abhishek Bhattacharya

Security of multimedia data is a major concern due to its widespread transmission over various communication channels. Hence design and study of good image encryption schemes has become a major research topic. During the last few decades, there has been a increasing in chaos-based cryptography. This paper proposes an attack on a recently proposed chaos based image encryption scheme. The cryptosystem under study proceed by first shuffling the original image to disturb the arrangement of pixels by applying a chaotic map several times. Second, a keystream is generated using Chen's chaotic system to mix it with the shuffled pixels to finally obtain the cipher image. A chosen ciphertext attack can be done to recover the system without any knowledge of the key. It simply demands two pairs of plaintext-ciphertext to completely break the cryptosystem.