Privacy Preserving Learning

The command line argument --privacy_activation implements aggregated learning by saving only those features that have seen a minimum threshold of users.

Motivation:

• In many real-world scenarios, the recommender cannot use the feature preferences of a user directly for learning due to privacy constraints.
• However, the recommender can learn from aggregated data which would uphold the privacy of the user.

Methodology:

• For each feature, a 32-bit vector is defined.
• We calculate a 5-bit hash of the tag of the example.
• For each feature weight updated by a non-zero value, we use the 5-bit hash to look up a bit in the 32-bit vector and set it to 1.
• When saving the weights into a file, we calculate the number of bits set to 1 for a feature. If it is greater than the threshold, the weights for that feature are saved.

Threshold:

• The default value of the threshold is 10.
• The number of trials for the first occurrence of m bits being flipped out of k bits which is a geometric distribution.
• Therefore, the probability of the next bit being flipped given n bits are already flipped is (k-n)/k.
• The expectation of a geometric distribution is 1/p, hence the expected waiting time until m bits are flipped is a summation from 0 to m-1 of the expectation k/(k-n).
• On calculation, the expected waiting time for flipping 10 bits out of 32 was 11.76, while it was 22.21 for flipping 10 bits out of 11.
• This implied that at least 12 unique users are needed in expectation to flip 10 bits out of 32 bits.

Implementation:

--privacy_activation : To activate the feature

--privacy_activation_threshold arg (=10) : To set the threshold

Future Work:

• Implement the feature for save_resume.
• Work on aggregations in the online setting.

Credits:

• Thanks to Olga Vrousgou and Pavithra Srinath for their mentorship.
• To further view the results for the empirical analysis of the feature, refer to : Empirical Analysis of Privacy Preserving Learning - RLOSF'21