Applying a deep learning algorithm to real financial trading records can effectively establish a profitable trading strategy, but leaking of private information during the process is a concern. To solve this problem, previous studies have proposed a privatized data “release” mechanism, but this mechanism is unable to truly protect privacy. To preserve privacy while still being able to learn useful information, we proposed a privacy-preserving generative adversarial imitation network (PPGAIN), a novel deep learning algorithm that imitates behaviors of general financial investors while preserving individual privacy. The core concept of PPGAIN is to provide class-ambiguous outputs to preserve privacy while generating sequences of trading behavior that are similar to a real investor's behavior. We used an auxiliary classifier as an adversary to ensure that the generated results are class ambiguous. We also proposed weakness replay to enforce the training of a discriminator. The PPGAIN model was trained in two stages, in the first stage, the model was trained to be realistic. In the second stage, it was trained against the auxiliary classifier to ensure that the outputs were highly realistic but could not be easily classified as an imitation from any specific individual to preserve privacy. Using real and synthetic trading records in TAIEX, we proved that our proposed method (PPGAIN) can achieve a generation quality similar to that obtained using conditional variational autoencoder models and demonstrated that PPGAIN provides better class ambiguity to preserve individual privacy than privacy-preserving adversarial network.
|Number of pages||16|
|State||Published - 21 Aug 2022|
- Deep learning
- Generative adversarial network
- Imitation learning