Differentially Private Stochastic Gradient Descent (DPSGD) is widely used to train deep neural networks with formal privacy guarantees. However, the addition of differential privacy (DP) often degrades model accuracy by introducing both noise and bias. Existing techniques typically address only one of these issues, as reducing DP noise can exacerbate clipping bias and vice-versa. In this paper, we propose a novel method, DP-PMLF, which integrates per-sample momentum with a low-pass filtering strategy to simultaneously mitigate DP noise and clipping bias. Our approach uses per-sample momentum to smooth gradient estimates prior to clipping, thereby reducing sampling variance. It further employs a post-processing low-pass filter to attenuate high-frequency DP noise without consuming additional privacy budget. We provide a theoretical analysis demonstrating an improved convergence rate under rigorous DP guarantees, and our empirical evaluations reveal that DP-PMLF significantly enhances the privacy-utility trade-off compared to several state-of-the-art DPSGD variants.

Software applications, while being an integral part of the modern world, pose significant threats to end-user privacy. Thus, computer professionals require knowledge and skills to develop privacy-aware software. However, undergraduate computing degree programs often lack privacy-focused curricula that can cultivate this ability in the future workforce. Therefore, we designed a privacy curriculum informed by the common challenges that computing professionals face when developing privacy-embedded software. It guides students in realising the need for privacy, identifying privacy protection mechanisms and programming Privacy Enhancing Technologies (PETs). We piloted the curriculum for third-year Computer Science undergraduates at the University of Auckland, New Zealand. The curriculum was evaluated using course assessments and surveys conducted before and after the lessons. Overall, the students improved their understanding of privacy, especially technical aspects. Most of them valued the applied learning experience of the programming lessons yet showed distinct views on task completion difficulty and motivation to do programming. Students recognised that privacy should be integral to their skill set by confirming the importance and relevance of the lessons. However, their perceived responsibility in privacy protection varied depending on their intention to take proactive measures. Based on the results, the paper suggests improvements to the proposed curriculum.

The rapid advancement of Artificial Intelligence (AI) has transformed various industries, leading to the widespread distribution of AI models and data across intelligent systems. As modern data driven services increasingly integrate distributed knowledge entities, decentralized learning has become a prevalent approach to training AI models. However, this collaborative learning paradigm introduces significant security vulnerabilities and privacy challenges. This paper presents a comprehensive systematic review on private knowledge sharing in distributed learning, analyzing key knowledge components utilized in leading distributed learning architectures. We identify critical vulnerabilities associated with these components and examine defensive strategies to safeguard privacy while mitigating potential adversarial threats. Additionally, we highlight key limitations in knowledge sharing in distributed learning and propose future research directions to enhance security and efficiency in decentralized AI systems.

Deep learning-based linkage of records across different databases is becoming increasingly useful in data integration and mining applications to discover new insights from multiple data sources. However, due to privacy and confidentiality concerns, organisations often are unwilling or allowed to share their sensitive data with any external parties, thus making it challenging to build/train deep learning models for record linkage across different organisations' databases. To overcome this limitation, we propose the first deep learning-based multi-party privacy-preserving record linkage (PPRL) protocol that can be used to link sensitive databases held by multiple different organisations. In our approach, each database owner first trains a local deep learning model, which is then uploaded to a secure environment and securely aggregated to create a global model. The global model is then used by a linkage unit to distinguish unlabelled record pairs as matches and non-matches. We utilise differential privacy to achieve provable privacy protection against re-identification attacks. We evaluate the linkage quality and scalability of our approach using several large real-world databases, showing that it can achieve high linkage quality while providing sufficient privacy protection against existing attacks.

The linkage of records to identify common entities across multiple data sources has gained increasing interest over the last few decades. In the absence of unique entity identifiers, quasi-identifying attributes such as personal names and addresses are generally used to link records. Due to privacy concerns that arise when such sensitive information is used, privacy-preserving record linkage (PPRL) methods have been proposed to link records without revealing any sensitive or confidential information about these records. Popular PPRL methods such as Bloom filter encoding, however, are known to be susceptible to various privacy attacks. Therefore, a systematic analysis of the privacy risks associated with sensitive databases as well as PPRL methods used in linkage projects is of great importance. In this article we present a novel framework to assess the vulnerabilities of sensitive databases and existing PPRL encoding methods. We discuss five types of vulnerabilities: frequency, length, co-occurrence, similarity, and similarity neighborhood, of both plaintext and encoded values that an adversary can exploit in order to reidentify sensitive plaintext values from encoded data. In an experimental evaluation we assess the vulnerabilities of two databases using five existing PPRL encoding methods. This evaluation shows that our proposed framework can be used in real-world linkage applications to assess the vulnerabilities associated with sensitive databases to be linked, as well as with PPRL encoding methods.

Entity resolution (ER) is the process of linking records that refer to the same entity. Traditionally, this process compares attribute values of records to calculate similarities and then classifies pairs of records as referring to the same entity or not based on these similarities. Recently developed graph-based ER approaches combine relationships between records with attribute similarities to improve linkage quality. Most of these approaches only consider databases containing basic entities that have static attribute values and static relationships, such as publications in bibliographic databases. In contrast, temporal record linkage addresses the problem where attribute values of entities can change over time. However, neither existing graph-based ER nor temporal record linkage can achieve high linkage quality on databases with complex entities , where an entity (such as a person) can change its attribute values over time while having different relationships with other entities at different points in time. In this article, we propose an unsupervised graph-based ER framework that is aimed at linking records of complex entities. Our framework provides five key contributions. First, we propagate positive evidence encountered when linking records to use in subsequent links by propagating attribute values that have changed. Second, we employ negative evidence by applying temporal and link constraints to restrict which candidate record pairs to consider for linking. Third, we leverage the ambiguity of attribute values to disambiguate similar records that, however, belong to different entities. Fourth, we adaptively exploit the structure of relationships to link records that have different relationships. Fifth, using graph measures, we refine matched clusters of records by removing likely wrong links between records. We conduct extensive experiments on seven real-world datasets from different domains showing that on average our unsupervised graph-based ER framework can improve precision by up to 25% and recall by up to 29% compared to several state-of-the-art ER techniques.

Privacy-preserving record linkage (PPRL) facilitates the matching of records that correspond to the same real-world entities across different databases while preserving the privacy of the individuals in these databases. A Bloom filter (BF) is a space efficient probabilistic data structure that is becoming popular in PPRL as an efficient privacy technique to encode sensitive information in records while still enabling approximate similarity computations between attribute values. However, BF encoding is susceptible to privacy attacks which can re-identify the values that are being encoded. In this paper we propose two novel techniques that can be applied on BF encoding to improve privacy against attacks. Our techniques use neighbouring bits in a BF to generate new bit values. An empirical study on large real databases shows that our techniques provide high security against privacy attacks, and achieve better similarity computation accuracy and linkage quality compared to other privacy improvements that can be applied on BF encoding.

To facilitate advanced analytics, data science projects increasingly require records about individuals to be linked across databases. Generally no unique entity identifiers are available in the databases to be linked, and therefore quasi-identifiers such as names, addresses, and dates of birth are used to link records. The process of linking records without revealing any sensitive or confidential information about the entities represented by these records is known as privacy-preserving record linkage (PPRL). Various encoding and encryption based PPRL methods have been developed in the past two decades. Most existing PPRL methods calculate approximate similarities between records because errors and variations can occur in quasi-identifying attribute values. Even though being used in real-world linkage applications, certain PPRL methods, such as popular Bloom filter encoding, have shown to be vulnerable to cryptanalysis attacks. In this paper we present a novel attack on PPRL methods that exploits the approximate similarities calculated between encoded records. Our attack matches nodes in a similarity graph generated from an encoded database with a corresponding similarity graph generated from a plain-text database to re-identify sensitive values. Our attack is not limited to any specific PPRL method, and in an experimental evaluation we apply it on three PPRL encoding methods using three different databases. This evaluation shows that our attack can successfully re-identify sensitive values from these encodings with high accuracy where no previous attack on PPRL would have been successful.

Privacy-preserving record linkage (PPRL) is the process of identifying records that correspond to the same entities across several databases without revealing any sensitive information about these entities. One popular PPRL technique is Bloom filter (BF) encoding, with first applications of BF based PPRL now being employed in real-world linkage applications. Here we present a cryptanalysis attack that can re-identify attribute values encoded in BFs. Our method applies maximal frequent itemset mining on a BF database to first identify sets of frequently co-occurring bit positions that correspond to encoded frequent q-grams (character substrings extracted from plain-text values). Using a language model, we then identify additional q-grams by applying pattern mining on subsets of BFs that encode a previously identified frequent q-gram. Experiments on a real database show that our attack can successfully re-identify sensitive values even when each BF in a database is unique.

Being able to identify records that correspond to the same entity across diverse databases is an increasingly important step in many data analytics projects. Research into privacy-preserving record linkage (PPRL) aims to develop techniques that can link records across databases such that besides the record pairs classified as matches no sensitive information about the entities in these databases is revealed. A popular technique used in PPRL is to encode sensitive values into Bloom filters (bit vectors), which has the advantage of allowing approximate matching using character q-grams. PPRL based on Bloom filter encoding has been shown to be accurate and scalable to large databases, and is thus now being used in real-world PPRL systems in Australia, Canada, and the UK. However, recent studies have shown that Bloom filters used for PPRL are vulnerable to cryptanalysis attacks that can re-identify some of the sensitive values encoded in these Bloom filters. While previous such attack methods were slow and required knowledge of various encoding parameters, we present a novel efficient attack which exploits how attribute values are encoded into Bloom filters. Our attack method does not require knowledge of the encoding function or its parameter settings used. It is able to correctly re-identify with high precision q-grams that could not have been hashed to certain Bloom filter bit positions, and using these re-identified q-grams it can then re-identify attribute values with high precision. Our method is significantly faster than earlier PPRL cryptanalysis attacks, and in our experimental evaluation, it is able to successfully re-identify attribute values from large real-world databases in a few minutes.

Record linkage (RL) is the process of identifying matching records from different databases that refer to the same entity. It is common that the attribute values of records that belong to the same entity do evolve over time, for example people can change their surname or address. Therefore, to identify the records that refer to the same entity over time, RL should make use of temporal information such as the time-stamp of when a record was created and/or update last. However, if RL needs to be conducted on information about people, due to privacy and confidentiality concerns organizations are often not willing or allowed to share sensitive data in their databases, such as personal medical records, or location and financial details, with other organizations. This paper is the first to propose a privacy-preserving temporal record linkage (PPTRL) protocol that can link records across different databases while ensuring the privacy of the sensitive data in these databases. We propose a novel protocol based on Bloom filter encoding which incorporates the temporal information available in records during the linkage process. Our approach uses homomorphic encryption to securely calculate the probabilities of entities changing attribute values in their records over a period of time. Based on these probabilities we generate a set of masking Bloom filters to adjust the similarities between record pairs. We provide a theoretical analysis of the complexity and privacy of our technique and conduct an empirical study on large real databases containing several millions of records. The experimental results show that our approach can achieve better linkage quality compared to non-temporal PPRL while providing privacy to individuals in the databases that are being linked.

Record linkage (RL) is a task in data integration that aims to identify matching records that refer to the same entity from different databases. When records from more than two databases are to be linked RL is significantly challenged by the intrinsic exponential growth in the number of potential record comparisons to be conducted. We propose a scalable meta blocking protocol to be used for Multi-Database RL (MDRL) to significantly reduce the complexity of the matching (comparison and classification) phase. Our approach uses a graph structure to schedule the comparison of pairs of blocks with the aim of minimizing the number of repeated and superfluous comparisons between records. We provide an analysis of our approach and conduct an empirical study on large real-world databases.