Quantamination: Dynamic quantization leaks user data within the same batch

The researchers found that dynamic quantization forms a side channel at runtime, leaking other users' data that is in the same batch as the attacker's input, and they call this phenomenon Quantamination. Dynamic quantization adjusts quantization parameters at runtime based on actual inputs; the paper states that if implemented or configured improperly, an attacker can exploit batch boundaries to obtain information about other inputs. The paper states that at least 4 of the most popular machine learning frameworks currently adopt by default or can use configurations that leak data across batches, which in theory could allow an attacker to partially or even fully recover the input data of other users in the same batch.