Privacy Matters: Unlock the Power of AI Without Compromising on Privacy

The Need

Artificial intelligence (AI) is rapidly evolving and has the potential to transform various industries and businesses. AI requires data, often in copious amounts, to function efficiently. The more granular the data, the better the algorithms can learn and make accurate predictions. Data collection has proliferated in recent years. Companies have come to value data as an asset and learned how to monetize it. Like any valuable asset, bad actors have an incentive to steal it. More than an economic loss, stolen data can become a liability for the company. Embarrassing data breaches and costly cyber-attacks require re-thinking how to create value with data while maintaining privacy.
Anonymized, pseudonoymized and synthetic data have gained prominence as approaches for dealing with data risk, not only safeguarding the privacy of individuals but encouraging more data subjects to willingly provide their details, thus helping to train AI. Several techniques may be employed to protect the privacy of individuals within data sets. Traditional approaches focus on data masking or content obfuscation through encryption, hash function and tokenization.

However, not all anonymization techniques are equally effective. The risk of re-identification remains. Companies need a reliable methodology to assess the adequacy of their privacy risk management measures. Advanced anonymization techniques, such as differential privacy, improve upon traditional masking methods by adding noise to the data. Combining such methods make it highly challenging to re-identify data subjects, while preserving the underlying distributions and patterns required to properly train a machine learning model.

Safely incorporating AI into processes, products and services is high on company agendas, and anonymization is a crucial component of that strategy. Whereas the importance and necessity of such objectives may be clear, how to achieve them is often less so.

Our solution: De-Identify

Deloitte's AI data anonymization and effectiveness assessment tool "De-Identify" provides a powerful, yet easy-to-use workflow for anonymizing both categorical and numerical tabular data. It enables users to both anonymize their data and assess the degree to which anonymization efforts successfully prevent re-identification.

De-Identify offers a wide range of methods for anonymization, including standard masking strategies such as data substitution, suppression, and generalization. It provides differential privacy anonymization by applying randomization using Laplace or exponential noise. De-Identify also enables deterministic masking and the preservation of correlation, facilitating the anonymization of data while preventing re-identification with minimal loss of expressiveness.

Furthermore, De-Identify provides multiple advanced privacy metrics to evaluate the level of anonymization and reduce risk of re-identification without requiring the original, adding to the tool's versatility in protecting data privacy.

Advantages/Benefits

• Examines privacy from multiple perspectives by assembling a collection of privacy indicators
• Uses both masking and differential privacy anonymization strategies
• Highly accessible to business users via the graphical user interface and straightforward, intuitive workflow
• Provides a means to generate synthetic data using anonymization techniques

Example Use Cases

• Sanity checking anonymization efforts by providing immediate feedback
• AI model validation for adherence to trustworthy AI principles
• Helping fulfill conformity to privacy compliance laws GDPR and AI Act
• Generation of representative synthetic data

Back to main page