Every time you use an app, visit a healthcare professional, or shop online, you’re creating data. This information fuels the artificial intelligence (AI) systems that help businesses enhance their offerings, enable doctors to identify illnesses sooner, and assist governments in making informed choices.

As AI evolves and increasingly depends on personal data, worries about how this information is utilized and protected are escalating. The central question emerging from this concern is whether we can continue to enjoy the benefits of advanced technology without sacrificing our privacy.

Sonakshi Garg, a doctoral candidate at Umeå University, believes we can have both. In her thesis titled “Bridging AI and Privacy: Solutions for High-Dimensional Data and Foundation Models,” Garg proposes innovative solutions designed to ensure that AI can respect personal data while remaining intelligent. She refers to this as the “privacy paradox”: Should we prioritize advanced AI or strong privacy?

“We shouldn’t have to choose between the two; we can have both,” she maintains.

To tackle this dilemma, Garg employs manifold learning to simplify complicated data while keeping its important structure intact. “Think of it like unfolding a wrinkled map while preserving the roads and landmarks—this is what manifold learning does with intricate datasets,” she explains.

Furthermore, she introduces a blended privacy model that leverages the best aspects of two existing methods. This allows users to have more control over what information is safeguarded while retaining more of the data’s utility. “It generates highly realistic ‘fake’ data that mimics real information without revealing anyone’s identity. This enables researchers and developers to train AI systems securely, without using sensitive data,” Garg elaborates.

Finally, she addresses the privacy risks linked to large AI models like GPT and BERT, which can unintentionally “memorize” confidential information. Her technique condenses these models, making them smaller and more efficient while adding extra privacy safeguards, allowing them to function securely even on devices like smartphones. Most importantly, Garg’s findings empower everyday users.

“It shows that individuals can enjoy tailored services and smart technologies without losing control over their personal lives. Privacy is not an obstacle to advancement—it’s the basis for creating better, more reliable AI,” she states.

As technology becomes more embedded in our daily routines, Sonakshi Garg’s research offers a vital framework for a future where AI and privacy can coexist harmoniously.

“My research serves as a critical reminder that intelligent innovation should never undermine human dignity—and with the correct tools, it doesn’t have to,” Sonakshi emphasizes.

This thesis explores the growing tension between the immense power of AI and the necessity to safeguard personal privacy in an era dominated by complex data. It identifies shortcomings in current privacy methods like k-anonymity and differential privacy when applied to intricate datasets, proposing enhanced solutions through manifold learning, synthetic data creation, and privacy-upholding model compression.

The study offers advanced, scalable methods that boost both the utility of data and privacy. Overall, the thesis presents a well-rounded strategy for building ethical, privacy-conscious AI systems suitable for real-world use.