Few-Shot Learning AMC

Employs architectures like Prototypical Networks that learn an embedding space where I/Q samples cluster tightly around a prototype representation for each modulation class. Classification of a new query sample is performed by finding the nearest class prototype in this learned metric space, requiring only a few support examples to define a new class centroid without retraining the entire network.

Utilizes Model-Agnostic Meta-Learning (MAML) to train an initial neural network weight configuration that is explicitly optimized for rapid fine-tuning. The inner loop performs a small number of gradient steps on the few available support samples, while the outer loop meta-updates the initialization to ensure that these few steps produce a highly generalizable classifier for novel modulation schemes.

Employs a hallucinator network that learns to generate additional synthetic I/Q samples for novel modulation classes from the few available real examples. The hallucinator is trained to produce plausible variations in phase noise, frequency offset, and fading conditions, effectively expanding the support set and preventing the classifier from overfitting to the limited real-world captures of a new threat signal.

Leverages the entire query set during inference to propagate labels from the few labeled support examples to the unlabeled queries. This approach constructs a graph where nodes represent I/Q samples and edges represent feature-space similarity, using label propagation algorithms to jointly classify the entire batch. This is particularly effective when a burst of unknown signals is intercepted simultaneously.

Uses cross-attention mechanisms to compute a similarity score between each support sample and a query sample. Instead of a single prototype, the model learns to focus on the most relevant support examples for each query, allowing it to handle intra-class variance in modulation parameters like symbol rate or pulse shaping. This provides robust matching even when the few available examples are noisy or imperfect.

Integrates elastic weight consolidation (EWC) or episodic memory replay to prevent the model from overwriting its ability to recognize previously learned modulation schemes when rapidly adapting to new ones. The algorithm identifies and protects critical weights important for legacy classes, ensuring the threat library expands cumulatively without degrading performance on existing signal types.

A self-supervised training paradigm that learns robust, discriminative I/Q embeddings without requiring labeled data. The model pulls augmented views of the same signal sample together in embedding space while pushing apart views from different samples. Key augmentations include:

• Phase rotation and frequency shifting
• Additive Gaussian noise at varying SNR levels
• Time cropping and amplitude scaling This produces a representation space where novel modulation classes naturally form distinct clusters, enabling few-shot classification via simple prototypical networks.

A metric-based few-shot learning architecture that computes a prototype vector for each modulation class by averaging the embeddings of the few available support examples. Classification of a query I/Q sample is performed by finding the nearest prototype in embedding space using Euclidean distance. This non-parametric approach excels in open-set recognition scenarios, as unknown modulation types naturally produce embeddings far from all known prototypes, enabling reliable out-of-distribution detection without retraining.

An optimization-based meta-learning algorithm that trains a model to find an internal representation highly sensitive to fine-tuning. During meta-training, the model is exposed to many few-shot tasks across diverse modulation families. The outer loop optimizes for initial parameters that can adapt to a new modulation type in 1-5 gradient steps. This is particularly effective for rapidly updating threat libraries in electronic warfare scenarios where new adversary waveforms appear with minimal intercepts.

Critical techniques that artificially expand limited training sets to prevent overfitting when only a handful of labeled I/Q samples are available:

• Channel simulation: Apply Rayleigh fading, Doppler shift, and multipath profiles
• Hardware impairment injection: Add carrier frequency offset (CFO), phase noise, and I/Q imbalance
• Mixup: Linearly interpolate between pairs of I/Q samples and their labels
• Noise augmentation: Train across a wide SNR range (-20dB to +30dB) to build SNR-invariant features

A classification paradigm where the AMC model must not only identify known modulation schemes but also detect and reject unknown types not seen during training. Few-shot learning naturally supports this through:

• Distance thresholding in embedding space
• Extreme value theory to model the probability that a sample belongs to any known class
• Energy-based models that assign low confidence scores to out-of-distribution signals This is essential for electronic warfare where adversaries continuously deploy novel waveforms.