Abstract Large Language Models (LLMs) have demonstrated exceptional capabilities in language understanding and generation, but their large-scale architecture poses significant challenges in deployment and inference, such as increased computational demands and slower processing times. While various techniques like model pruning, knowledge distillation, and quantization have been developed to compress LLMs, they often result in task-specific compression, limiting the model’s versatility. Additionally, LLMs face privacy risks due to their potential to memorize and reproduce sensitive training data, raising concerns when deployed in real-world applications. To address these challenges, we propose a novel methodology PrunePrivyTune that combines efficient model compression with privacy preserving fine-tuning. Our approach leverages pairwise cosine similarity to identify redundant layers in transformer models, enabling structural pruning that reduces model size without compromising performance. After pruning, we apply Low-Rank Adaptation (LoRA) with DPSGD to fine-tune the model. This ensures that fine-tuning process is both efficient and privacy-preserving, outperforming training and preventing the model from memorizing sensitive data. Later on, we generated synthetic data using the fine-tuned model and subsequently conducted a training data extraction attack to assess the model’s privacy vulnerabilities, in terms of perplexity and BERTScore. Our framework demonstrates that the proposed methodology effectively reduces the inference time through model compression and pruning compliments privacy, followed by private fine-tuning. Additionally, our privacy risk assessment indicates that integrating DP successfully mitigates the risk of the model’s memorization. This approach upholds strong privacy guarantees, making it highly suitable for real-time applications and deployment in sensitive domains where data confidentiality is paramount.

While anonymization techniques have improved greatly in allowing data to be used again, it is still really hard to get useful information from anonymized data without risking people’s privacy. Conventional approaches such as k-Anonymity and Differential Privacy have limitations in preserving data utility and privacy simultaneously, particularly in high-dimensional spaces with manifold structures. We address this challenge by focusing on anonymizing data existing within high-dimensional spaces possessing manifold structures. To tackle these issues, we propose and implement a hybrid anonymization scheme termed as the (β, k, b)-anonymization method that combines elements of both differential privacy and k-anonymity. This approach aims to produce high-quality anonymized data that closely resembles real data in terms of knowledge extraction while safeguarding privacy. The Fréchet mean, an operation applicable in metric spaces and meaningful in the manifold setting, serves as a key aspect of our approach. It provides insight into the geometry of data points within high-dimensional spaces. Our goal is to anonymize this Fréchet mean using our proposed approach and minimize the distance between the original and anonymized Fréchet mean to achieve data privacy without significant loss of information. Additionally, we introduce a novel Fréchet mean clustering model designed to enhance the clustering process for high-dimensional spaces. Through theoretical analysis and practical experiments, we demonstrate that our approach outperforms traditional privacy models both in terms of preserving data utility and privacy. This research contributes to advancing privacy-preserving techniques for complex and non-linear data structures, ensuring a balance between data utility and privacy protection.

In this modern world of digitalization, abundant amount of data is being generated. This often leads to data of high dimension, making data points far-away from each other. Such data may contain confidential information and must be protected from disclosure. Preserving privacy of this high-dimensional data is still a challenging problem. This paper aims to provide a privacy preserving model to anonymize high-dimensional data maintaining the manifold structure of the data. Manifold Learning hypothesize that real-world data lie on a low-dimensional manifold embedded in a higher-dimensional space. This paper proposes a novel approach that uses geodesic distance in manifold learning methods such as ISOMAP and LLE to preserve the manifold structure on low-dimensional embedding. Later on, anonymization of such sensitive data is achieved by M-MDAV, the manifold version of MDAV using geodesic distance. MDAV is a micro-aggregation privacy model. Finally, to evaluate the efficiency of the prop osed approach machine learning classification is performed on the anonymized lower-embedding. To emphasize the importance of geodesic-manifold learning, we compared our approach with a baseline method in which we try to anonymise high-dimensional data directly without reducing it onto a lower-dimensional space. We evaluate the proposed approach over natural and synthetic data such as tabular, image and textual data sets, and then empirically evaluate the performance of the proposed approach using different evaluation metrics viz. accuracy, precision, recall and K-Stress. We show that our proposed approach is providing accuracy up to 99% and thus, provides a novel contribution of analysing the effects of K-anonymity in manifold learning.