Data-Free Quantization (DFQ)

DFQ exploits the batch normalization layers frozen inside a pre-trained model to reconstruct the channel-wise mean and variance of the lost training data. By sampling from these stored Gaussian distributions, the algorithm generates synthetic images or IQ samples that statistically mimic the original input domain. This eliminates the need to store sensitive RF training datasets while providing sufficient calibration data to minimize the Kullback-Leibler divergence between the floating-point and quantized activation ranges.

Advanced DFQ methods train a lightweight generative model to produce realistic inputs without accessing real data. The generator is optimized adversarially to maximize the discrepancy between the full-precision and quantized model outputs, forcing the creation of hard boundary samples. This approach is critical for automatic modulation classification where the original signal captures are classified or operationally sensitive, allowing calibration on synthetic constellations that preserve phase and magnitude fidelity.

ZSQ is a strict subset of DFQ where absolutely no data samples—real or synthetic—are used for calibration. Instead, the quantization parameters are derived purely from the mathematical properties of the weight tensors and activation functions. Techniques include weight equalization to balance per-channel ranges and bias correction to absorb the systematic error introduced by rounding. This is essential for deploying neural receivers on secure FPGA fabric where running a generator is infeasible.

To prevent accuracy collapse at low bit-widths, DFQ frameworks use boundary-seeking generative adversarial networks (GANs). The generator creates inputs that lie exactly on the decision boundaries of the classifier, which are the most susceptible to quantization noise. By calibrating the INT8 scaling factors on these adversarial examples, the quantized model maintains robust classification margins for radio frequency fingerprinting applications where subtle hardware impairments must survive precision reduction.

Some DFQ pipelines optimize random noise directly against the frozen model's feature statistics, a process akin to DeepDream. By maximizing the activation of specific feature maps, the algorithm visualizes and reconstructs the class-conditional input distributions the network expects. For IQ sample processing, this recovers the complex Gaussian noise structure of the baseband signal, enabling accurate calibration of the straight-through estimator (STE) without storing raw in-phase and quadrature recordings.

DFQ must decide the granularity of quantization parameters without a validation set. Per-tensor quantization uses a single scale for an entire weight matrix, while per-channel quantization assigns a distinct scale to each output channel. DFQ algorithms analyze the weight distribution patterns to automatically select the optimal granularity. This is vital for depthwise separable convolutions in TinyML models, where per-channel quantization prevents catastrophic accuracy loss in the depthwise layers.