In today’s digital age, the importance of safeguarding personal data from various threats, such as malicious adversaries, social networking sites, and privacy-invading employees, has become paramount. This heightened awareness of data collection and protection regulations has spurred a growing interest in privacy-enhancing technologies (PETs).
Figure 1 describes the evolution of the data privacy domain and corresponding developments.
Source: Rectification of Syntactic and Semantic Privacy Mechanisms
A typical, deployed AI-enabled system concerns two types of personnel, system owner and system user, and both are subject to data protection risks.
System Owners: The system owner is responsible for safeguarding both the AI model and the data used to train it. Protecting the AI model involves keeping its characteristics private and maintaining its integrity against inputs designed to deceive or tamper with the model. Additionally, the training data must be protected to prevent users from uncovering details about the data through the model. This also includes preventing any data leaks by those who develop or maintain the system.
System Users: On the other hand, system users need to ensure that the data they send to and process through the AI-enabled system remains secure. When users request inferences over private data, they must be confident that their personal information will not be exposed, even to the system owner. AI systems often learn user behaviours and characteristics over time, making it crucial for users to trust that their data will not be misused or repurposed without their consent.
PETs are vital tools for mitigating vulnerabilities faced by both system owners and users. These technologies encompass a range of techniques designed to process data while protecting individuals from unwanted disclosures. Key PETs include:
- Differential Privacy: Ensures that statistical analyses do not reveal individual data points. It is based on the randomised injection of noise.
- Homomorphic Encryption: Allows computations on encrypted data without needing to decrypt it.
- Secure Multiparty Computation: Enables parties to jointly compute a function over their inputs while keeping those inputs private.
In the context of AI, PETs are particularly relevant through privacy-preserving machine learning (PPML). PPML integrates PETs at various stages of the machine learning process to train models over encrypted data, anonymise the training process, and protect outputs using differential privacy. This combination of technologies helps defend against privacy attacks and ensures that both system owners and users can trust the AI-enabled system.
Figure 2 sets out a selection of emerging and legacy PETs.
PETs play a crucial role in safeguarding privacy and data protection through various innovative approaches. These technologies can be broadly categorised into three groups:
1. Data-Altering Tools: These PETs aim to disrupt or break the connection between data and the individuals they are associated with. By altering the data itself, these tools ensure that personal information cannot be easily traced back to its source. Techniques such as anonymisation and pseudonymisation fall into this category, where identifiable information is removed or replaced to protect individual privacy.
2. Data-Hiding Tools: This group of PETs focuses on hiding or shielding data rather than altering it. Encryption is a prime example, as it changes the format of data to obscure it temporarily, ensuring that only authorised parties can access the original information. Other methods include secure multiparty computation and differential privacy, which protect data during processing and analysis without permanently altering it.
3. New Systems and Data Architectures: The third category encompasses PETs that introduce new systems and architectures for processing, managing, and storing data. These technologies often involve breaking apart data for computation or storage, ensuring that sensitive information is not concentrated in a single location. Additionally, management layers are implemented to track and audit the flow of information, providing transparency and control over data usage.
By leveraging these diverse PETs, organisations can enhance their data protection strategies, ensuring that personal information remains secure and private. Whether through altering, hiding, or rearchitecting data, PETs offer robust solutions to address the evolving challenges of data privacy.
Figure 3 describes a selection of PETs organised by category.
PETs are integral to the concept of ‘data protection by design,’ ensuring that both technical and organisational measures are in place to safeguard personal data. These technologies help implement data protection principles effectively and integrate necessary safeguards into data processing activities. By adopting PETs, organisations can demonstrate a ‘data protection by design and by default’ approach through several key practices:
- Complying with the Data Minimisation Principle: Ensuring that only the necessary information is processed for specific purposes.
- Providing an Appropriate Level of Security: Implementing robust security measures to protect data.
- Implementing Anonymisation or Pseudonymisation Solutions: Reducing the risk of identifying individuals from the data.
- Minimising the Risk of Data Breaches: Making personal information unintelligible to unauthorised individuals.
While PETs often involve processing personal information, it is crucial to ensure that such processing remains lawful, fair, and transparent. Conducting a case-by-case assessment, such as a data protection impact assessment (DPIA), helps identify risks to individuals and determine if PETs are appropriate to mitigate those risks. It is important to note that not all PETs result in effective anonymisation, and anonymisation can be achieved without using PETs.
Anonymisation is a practical technique for sharing data without compromising individual privacy. It involves removing or modifying data items that can directly identify individuals, balancing privacy and utility. There are two main mechanisms for preserving individual privacy in data analysis:
- Syntactic Mechanisms: These alter the dataset before release to ensure that any record is linked to more than one sensitive value. Examples include k-anonymity, ℓ-diversity, and t-closeness. Initially designed for tabular data, these models have faced challenges from contemporary privacy threats, prompting updates and improvements.
- Semantic Mechanisms: These limit the impact of individual values on the query or analysis output based on the dataset. Differential privacy (DP) is a leading semantic mechanism that offers strong privacy guarantees for dynamic scenarios like query-answering.
By leveraging these mechanisms, organisations can enhance their data protection strategies and ensure that personal information remains secure and private.
PETs represent the next step forward for cybersecurity. No longer can we only protect data at rest and in transit. We must also protect data in use, thus closing the final gap and providing true end-to-end security. When combined with AI, privacy-enhancing technologies are crucial tools for safeguarding data privacy, boosting security, and encouraging responsible data governance; they are not only for compliance or legal teams.
If the rise of AI-enabled systems is accompanied by the rise of PET-based protection, then the privacy risks associated with these systems can be greatly diminished. They are absolutely necessary for today’s data-driven environment because of the advantages they provide, such as improved privacy and accelerated innovation. PETs give businesses a variety of ways to safeguard customer information while maximising data use. Organisations may guarantee a secure and privacy-preserving data infrastructure by comprehending the various types of data and use cases, evaluating regulatory requirements, and choosing the appropriate PET for their business.
Resources
Privacy Enhancing Technologies – Key… | Mason Hayes Curran
Privacy-enhancing technologies (PETs) | ICO
Rectification of Syntactic and Semantic Privacy Mechanisms
Privacy-Enhancing Technologies for Artificial Intelligence-Enabled Systems
AI-Driven Privacy Enhancing Technologies (PETs) and Clarifying Misconceptions About It | LinkedIn