An Attack-Defense Game-Based Reinforcement Learning Privacy-Preserving Method Against Inference Attack in Double Auction Market

Donghe Li, Chunlin Hu, Qingyu Yang, Yuhao Ma, Feiye Zhang, Dou An

Abstract:

Auction mechanism, as a fair and efficient resource
allocation method, has been widely used in varieties trading
scenarios, such as advertising, crowdsensoring and spectrum.
However, in addition to obtaining higher profits and satisfaction,
the privacy concerns have attracted researchers’ attention. In this
paper, we mainly study the privacy preserving issue in the double
auction market against the indirect inference attack. Most of
the existing works apply differential privacy theory to defend
against the inference attack, but there exists two problems. First,
‘indistinguishability’ of differential privacy (DP) cannot prevent
the disclosure of continuous valuations in the auction market.
Second, the privacy-utility trade-off (PUT) in differential privacy
deployment has not been resolved. To this end, we proposed
an attack-defense game-based reinforcement learning privacy
preserving method to provide practically privacy protection in
double auction. First, the auctioneer acts as defender, adds noise
to the bidders’ valuations, and then acts as adversary to launch
inference attack. After that the auctioneer uses the attack results
and auction results as a reference to guide the next deployment.
The above process can be regarded as a Markov Decision Process
(MDP). The state is the valuations of each bidders under the
current steps. The action is the noise added to each bidders.
The reward is composed of privacy, utility and training speed,
in which attack success rate and social welfare are taken as
measures of privacy and utility, a delay penalty term is used to
reduce the training time. Utilizing the deep deterministic policy
gradient (DDPG) algorithm, we establish an actor-critic network
to solve the problem of MDP. Finally, we conducted extensive
evaluations to verify the performance of our proposed method.
The results show that compared with other existing DP-based double auction privacy preserving mechanisms, our method can
achieve better results in both privacy and utility. We can reduce
the attack success rate from nearly 100% to less than 20%, and
the utility deviation is less than 5%.