Denoising Autoencoder

The core mechanism involves applying a stochastic mapping to the clean input x to produce a corrupted version x̃. Common RF-specific corruptions include:

• Additive White Gaussian Noise (AWGN): Adding thermal noise to IQ samples
• Masking Noise: Randomly zeroing a fraction of IQ components
• Salt-and-Pepper Noise: Setting random IQ values to extreme amplitudes
• Channel Impairments: Simulating fading, phase offset, or non-linearity The model is trained to minimize the reconstruction loss L(x, g(f(x̃))), where f is the encoder and g is the decoder. This forces the hidden layer to learn a noise-invariant manifold of the signal space.

By removing noise rather than simply copying input, the DAE learns the underlying data manifold—the lower-dimensional subspace where clean signals reside. Key properties:

• The encoder maps corrupted signals back toward the manifold
• The reconstruction function locally estimates the score (gradient of the log-density) of the data distribution
• In RF applications, this captures the intrinsic modulation structure independent of instantaneous SNR This makes DAEs superior to standard autoencoders for pre-training on unlabeled spectrum captures where noise levels vary dramatically.

Individual DAEs can be stacked greedily layer by layer to form deep networks:

1. Train first DAE on corrupted raw IQ data to learn low-level features
2. Use its encoder output as input to a second DAE, learning mid-level representations
3. Repeat for desired depth
4. Fine-tune the entire stack end-to-end with a supervised objective This layer-wise pre-training was foundational in deep learning and remains effective for RF tasks with limited labeled data. The resulting deep network initializes in a region of weight space that captures hierarchical signal structure—from transient edges to modulation-specific patterns.

A trained DAE with Gaussian corruption noise implicitly learns the score function of the data distribution. The optimal denoising function satisfies: g(f(x̃)) - x̃ ∝ ∇ log p(x̃) This connects DAEs to Denoising Diffusion Probabilistic Models (DDPMs) and Score-Based Generative Models. In RF contexts, this means:

• The DAE can be interpreted as one step of a Langevin dynamics sampling chain
• The learned representations align with the energy landscape of legitimate signals
• This provides theoretical grounding for why DAEs generalize well to unseen interference patterns

A typical DAE pre-training workflow for modulation recognition:

1. Collect: Gather millions of unlabeled IQ captures from wideband spectrum monitoring
2. Corrupt: Apply variable AWGN, phase rotation, and DC offset to create input-target pairs
3. Architecture: Use a 1D convolutional encoder with strided convolutions and a symmetric transposed convolution decoder
4. Loss: Minimize Mean Squared Error (MSE) between clean and reconstructed IQ
5. Transfer: Discard decoder, attach a classification head, and fine-tune on a few hundred labeled modulation samples This approach routinely achieves 85-95% of fully-supervised performance using only 5-10% of the labeled data.

DAEs provide inherent resilience against physical layer adversarial perturbations:

• Evasion Jamming: The denoising process removes crafted interference designed to fool downstream classifiers
• Reactive Distortion: Non-linear amplifier effects are treated as a corruption domain the DAE learns to invert
• Defensive Pre-processing: A trained DAE can be deployed inline as a front-end filter before a legacy classifier without retraining the downstream model Empirical studies show DAE-based receivers maintain >90% classification accuracy under jamming-to-signal ratios where unprotected models drop below 30%.

The central, compressed representation of the input data within an autoencoder. By forcing the signal through a low-dimensional latent space, the network is structurally prevented from learning the identity function. In a denoising autoencoder, this bottleneck must capture only the essential, robust features of the clean signal while discarding the stochastic noise pattern. The dimensionality of this layer is a critical hyperparameter that controls the trade-off between compression and reconstruction fidelity.

The objective function that quantifies the difference between the original clean signal and the decoder's output. Common choices for RF IQ data include:

• Mean Squared Error (MSE): Penalizes large amplitude deviations.
• Mean Absolute Error (MAE): Less sensitive to outliers. The loss is computed against the clean target, not the corrupted input, which is the defining characteristic that forces the model to learn to subtract noise.

The stochastic mapping that transforms a clean input sample into its corrupted counterpart before feeding it to the encoder. In RF applications, this is physically meaningful:

• Additive White Gaussian Noise (AWGN): Simulates thermal receiver noise.
• Masking Noise: Randomly zeros out IQ samples to mimic packet loss or impulsive interference.
• Phase Noise: Applies random rotations to simulate local oscillator instability. The model must learn the statistical structure of the corruption to invert it.

A deep architecture formed by stacking multiple denoising autoencoders, where the hidden representation of one serves as the input to the next. This greedy layer-wise pre-training strategy was foundational in deep learning, enabling the unsupervised initialization of deep networks before supervised fine-tuning. Each layer learns to denoise its input, progressively building higher-level, more invariant representations of the underlying signal structure.

The theoretical framework underpinning denoising autoencoders. The clean RF signals are assumed to lie on a low-dimensional manifold embedded in the high-dimensional observation space. The corruption process stochastically pushes samples off this manifold. The denoising autoencoder learns a vector field that maps corrupted points back toward the nearest point on the clean data manifold, effectively learning the score (gradient of the log-probability density) of the data distribution.

A closely related self-supervised architecture that can be viewed as a specific, aggressive form of denoising. Instead of adding continuous noise, an MAE masks a large fraction of input patches (e.g., 75% of an IQ sequence) and tasks the decoder with reconstructing the missing segments. Unlike a traditional denoising autoencoder, the MAE's encoder only processes the visible, unmasked patches, making it highly compute-efficient for pre-training on long signal sequences.