How can agents work together but keep secrets
at the same time?
In learning and control, multi-agent systems function by having agents
exchange information. Over time, the information being shared
has included more and more user data. For example, self-driving cars share
location data to be routed to a destination and smart homes share power
usage data to ensure that power generation is sufficient. Even
seemingly benign user data can reveal sensitive details about
users' lives, and thus a paradox emerges: agents must share
information to accomplish some goal, though they want
to protect the same information that they are sharing. In
other words, agents must share sensitive data in a way that
protects it from its recipient while preserving its usefulness
to that recipient. Encryption does not do so because
it either reveals the plaintext to a recipient (i.e., reveals
the sensitive data itself) or shares only a ciphertext with a
recipient (which reveals nothing to them). Instead, a tunable
form of privacy is required to share sensitive data in
measured amounts.
Differential privacy has been used for this purpose in the
computer science literature, and it has been successful at
privatizing data of many types and in many settings.
Our work departs from much of the existing literature because
we study systems in learning and control with feedback.
This requires fundamental innovations to incorporate
differential privacy into data-driven systems and to
rigorously ensure that system performance is preserved.
We have developed such
implementations and analyses for problems throughout control, learning, and autonomy,
a sampling of which includes:
1. K. Yazdani, A. Jones, K. Leahy, and M.T. Hale,
"Differentially Private LQ Control," Under
review. [preprint].
2. M.T. Hale and M. Egerstedt, "Cloud-enabled
differentially private multi-agent optimization with
constraints,"
IEEE Transactions on Control of Network Systems,
vol. 5, no. 4, pp. 1693-1706, 2018. [link]
3. A. Jones, K. Leahy, and M.T. Hale, "Towards Differential Privacy for Symbolic Systems,"
2019 American Control Conference (ACC), 2019, pp. 372-377.
[paper][preprint]
4. P. Gohari, M.T. Hale, and U. Topcu,
"Privacy-Preserving Policy Synthesis in Markov Decision Processes,"
59th IEEE Conference on Decision and Control, pp. 6266-6271.
[preprint]
5. C. Hawkins and M.T. Hale,
"Differentially private formation control,"
59th IEEE Conference on Decision and Control, pp. 6260-6265.
[preprint]
