All-domain Anomaly Resolution Office U.S. Department of Defense (U) Case: “Eg...
Listener Anonymizer: Camouflaging Play Logs to Preserve User’s Demographic Anonymity (ISMIR 2018)
1. Listener Anonymizer:
Camouflaging Play Logs
to Preserve User’s Demographic Anonymity
Kosetsu Tsukuda, Satoru Fukayama, Masataka Goto
National Institute of Advanced Industrial Science and Technology (AIST), Japan
Sept. 26, 2018
2. I love music recommendation
Music recommendation can improve user’s music experience
3. To improve recommendation accuracy, it is beneficial to
predict user’s demographic attributes (age, gender, nationality)
A user’s demographics can be predicted with high accuracy
by using the user’s play log
𝑡𝑡
Your nationality is
Age Gender Nationality
4.13mean absolute error 77.01%accuracy 69.37%accuracy
T. Krismayer, M. Schedl, P. Knees, R. Rabiser
Prediction of User Demographics from Music Listening Habits
CBMI 2017
Play log
4. Technique to leverage play logs
for predicting users' demographic attributes
?
COUNTERBALANCE
5. Technique to leverage play logs
for predicting users' demographic attributes
Technique to camouflage play logs
for preserving users' demographic anonymity
COUNTERBALANCE
7. 𝑡𝑡
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
8. 𝑡𝑡
Listener Anonymizer
…
Compute
a probability distribution
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
9. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
10. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
11. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
…
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
12. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
Listener Anonymizer
Your nationality can be predicted
as French with a probability of 67%
Anonymize
…
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
13. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
…
0.21 0.04 0.77 0.89
0.48 0.92 0.25 0.33
0.82 0.29 0.46 0.86
𝑢𝑢1
𝑢𝑢2
𝑢𝑢𝑟𝑟
𝑠𝑠1 𝑠𝑠2 𝑠𝑠3 𝑠𝑠𝑚𝑚… …
Compute the effectiveness of each song
to anonymize her nationality
𝑢𝑢𝑖𝑖: user
𝑠𝑠𝑗𝑗: song
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
14. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
…
0.21 0.04 0.77 0.89
0.48 0.92 0.25 0.33
0.82 0.29 0.46 0.86
𝑢𝑢1
𝑢𝑢2
𝑢𝑢𝑟𝑟
𝑠𝑠1 𝑠𝑠2 𝑠𝑠3 𝑠𝑠𝑚𝑚… …
Compute the effectiveness of each song
to anonymize her nationality
Listener Anonymizer
Recommendations:
Play
𝑢𝑢𝑖𝑖: user
𝑠𝑠𝑗𝑗: song
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
15. 𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
…
0.21 0.04 0.77 0.89
0.48 0.92 0.25 0.33
0.82 0.29 0.46 0.86
𝑢𝑢1
𝑢𝑢2
𝑢𝑢𝑟𝑟
𝑠𝑠1 𝑠𝑠2 𝑠𝑠3 𝑠𝑠𝑚𝑚… …
Compute the effectiveness of each song
to anonymize her nationality
𝑢𝑢𝑖𝑖: user
𝑠𝑠𝑗𝑗: song
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
16. …
𝑡𝑡
Listener Anonymizer
Compute
a probability distribution
…
0.21 0.04 0.77 0.89
0.48 0.92 0.25 0.33
0.82 0.29 0.46 0.86
𝑢𝑢1
𝑢𝑢2
𝑢𝑢𝑟𝑟
𝑠𝑠1 𝑠𝑠2 𝑠𝑠3 𝑠𝑠𝑚𝑚… …
Compute the effectiveness of each song
to anonymize her nationality
Your nationality is … ??
𝑢𝑢𝑖𝑖: user
𝑠𝑠𝑗𝑗: song
Emma is a 22-year-old French female
She uses both an online music service and Listener Anonymizer
She concealed her nationality when she signed up to the service
18. Probability When a user plays 30 songs,
the distribution is strongly biased to Polish (the left most graph)
She can camouflage her play log by playing only three songs
recommended by Listener Anonymizer
19. I can enjoy music while preserving
my demographic anonymity!
Without Listener Anonymizer With Listener Anonymizer
It is important to show that
preserving users’ demographic anonymity is technically possible
Listener Anonymizer gives a choice to a user
Demographic
anonymity
High rec.
accuracy
I do not care about
my demographic anonymity!
20. Listener Anonymizer might degrade recommendation accuracy
We dared to propose this controversial approach
to raise privacy issues in the ISMIR community