Fighting Fire with Fire: New JKU AI Conceals Personal Information from AI Recommendation Systems

The JKU has developed a method to conceal individual, personal traits when searching for information but without compromising search and recommendation results.

Professor Markus Schedl
Professor Markus Schedl

Interested in finding out just which artists are currently topping the pop charts? Or would you like some suggestions to see a movie today? Modern search engines and recommendation systems apply artificial intelligence and deep learning methods, meaning you get a precise answer to your question but at the same time, the system can also predict your personal preferences with astonishing accuracy. For example, in regard to movies or music, your gender can be correctly inferred 72% of the time just from your user behavior. Researchers at the Johannes Kepler University Linz have recently developed a method designed to conceal your individual preferences but without impairing the search and recommendation results.

In an effort to prevent recommendation systems from predicting the user’s personal preferences, Prof. Markus Schedl and his colleagues at the JKU Institute for Computational Perception have developed their own deep learning architecture: Adversarial Variational Auto-Encoder with Multinomial Likelihood, in short Adv-MultVAE.

During this process, two networks are activated and while they work separately, they are also closely linked. One network solves the actual task, meaning it finds the music or movies that the user might be interested in. The second network makes the big difference as it tries to predict sensitive personal information (such as gender or nationality) as accurately as possible.

In contrast to conventional recommendation system algorithms, the JKU networks take on a different direction at that point. The working parameters are gradually adjusted so that while recommendations results remain the same, the system is less accurate when it comes to predicting the user’s personal attributes.

In the end, the networks experience an evolution that provides good search results but is weak when it comes to predicting personal information.

A Lower Rate of Predictive Accuracy
Researchers tested the new system using music and movie datasets. The prediction accuracy in regard to the user’s gender was reduced by 57% when it came to movies and 62% when it came to music; in terms of recommendation quality, the results remained the same.

In addition to protecting the user's personal information, this method also has another advantage. Markus Schedl explains: "Because our new recommendation systems no longer know the users’ exact genders, it suggests a broader spectrum of music and movies. This means that men also get recommendations for movies often considered to be ‘chick flicks’, such as Casablanca. As a result, our approach also counteracts stereotypes as well as creating ‘bubbles’ on the Internet." The JKU algorithm has yet to work better for women. Schedl added: "It is probably on account of simply having more information provided by men, or women’s user behavior is different than men." In this regard, there are plans to further develop the system and include a stronger focus on traits that are considered to be more feminine.

The Algorithm is Openly Available
The findings and the algorithm were recently published during the Top Information Retrieval Conference (ACM SIGIR) and are openly available. Schedl remarked: "Everything you need to implement it, such as the source code, datasets, etc., is available in our Github account. So, if you have a bit of expertise in the area, we invite you to test the algorithm yourself."

Publication: https://dl.acm.org/doi/10.1145/3477495.3531820, opens an external URL in a new window

Github Account: https://github.com/CPJKU/adv-multvae, opens an external URL in a new window