AC-LIA: The Labels Divulge Your Secrets and Privacy

The emergence of federated learning helps multiple participants to optimize model and process data, and it ensures that models are trained without sharing raw data, leading to more efficient and secure model training. However, there are many security challenges in federated learning, such as Label Inference Attack (LIA) in vertical federated learning. That is, the attacker obtains the private key label information data during the training process. Current label inference attacks exhibit a notably elevated false positive rate (FPR), which means they frequently forecast incorrect data labels, making the attack impractical. And previous research only focuses on the precision of the attacks without paying attention to the False Forecasting Rate (FPR). In this paper, we put forward a low-frequency-ratio, high-precision pre-calibration (AC-LIA) based label inference attack method. More specifically, we design a reference model which outputs a threshold value. Samples with a posteriori probability less than the threshold are then eliminated. Consequently, the model for malicious attacks is developed using a training dataset. with no removed samples and is characterized by a low false alarm rate. Experiments show that on the BCW dataset, the FPR of AC-LIA decreases from the original 45.70% to 9.23%, which is a significant decrease of about 5 times. The precision of the original LIA and AC-LIA is 88.67% and 88.93%, respectively. In other words, AC-LIA can effectively reduce the FPR while ensuring the precision of the original attack. Finally, countermeasures against AC-LIA are put forward.