Runs on CPU Too! FuseAMC: A New Model for Modulation Recognition Fusing Deep and Classical Methods

Automatic modulation classification (AMC) is a cornerstone of spectrum sensing and non-cooperative signal identification in cognitive radio. For a long time, two CPU-runnable approaches have each excelled in their own way: one is a compact convolutional neural network (CNN) that directly processes raw I/Q samples, and the other is a classical classifier based on higher-order cumulant features. But they have rarely been systematically compared under signal-to-noise ratio (SNR) conditions. A recent study reveals that the two are actually complementary: the CNN is clearly more robust at low SNR, while the cumulant classifier is more accurate at high SNR. Based on this, the researchers propose FuseAMC, a lightweight decision-fusion model that achieves performance beyond each of its individual baselines on the RadioML 2016.10a dataset, and the entire pipeline can be trained and run on just a single CPU core.

Where Does the Complementarity Come From

The study first compares the two baseline methods under controlled conditions. The CNN directly learns patterns in the I/Q time series, and in the low-SNR regime, when handcrafted statistical features approach random guessing because they are drowned out by noise, the CNN can still maintain considerable classification capability. Conversely, the cumulant classifier relies on higher-order statistics to distinguish dense constellations and performs better at high SNR, where higher-order moments can clearly separate signals. This complementary pattern—"deep models for low SNR, classical features for high SNR"—provides a direct motivation for fusion.

Higher-order cumulants are a classical statistical tool in signal processing, used to describe the extent to which a random process deviates from a Gaussian distribution. For modulated signals, second-order cumulants (variance) reflect power, fourth-order cumulants can distinguish shape differences among constellation diagrams, and sixth-order and above can further separate dense constellations (such as 16QAM and 64QAM). Their key advantage is that the higher-order cumulants of Gaussian noise are theoretically zero, so cumulant features have a natural suppression capability against additive white Gaussian noise. However, at low SNR, the variance of cumulants estimated from finite samples increases sharply, the feature values are drowned out by estimation noise, and classifier performance rapidly degrades to the level of random guessing. This is precisely the fundamental reason why cumulant methods are accurate at high SNR and fail at low SNR, and it also constitutes the physical basis for their complementarity with CNNs.

FuseAMC: Lightweight Decision Fusion

The core of FuseAMC is a logistic regression meta-combiner that takes the concatenated class posterior probabilities of two sub-models as input and is trained on a validation set disjoint from the test set. The two sub-models are a residual CNN with about 198k parameters and a gradient boosted tree (GBT) cumulant classifier. The entire model has about 0.2 million parameters, a footprint of only a few megabytes, and about 3 MFLOPs of computation per frame, so it requires no hardware acceleration and can complete training and inference on a single CPU core on general-purpose or embedded processors.

Decision fusion and feature fusion are two different multi-model ensemble strategies. Feature fusion concatenates features from different sources and feeds them into a single classifier, requiring the features to be semantically and scale-aligned; decision fusion occurs after each sub-model outputs independently and only combines the class posterior probabilities or decisions of each model, so the sub-models can be heterogeneous or even black boxes. The logistic regression meta-combiner adopted by FuseAMC is a typical stacking-style decision fusion: it uses the concatenated class posterior probabilities of the two sub-models as meta-features and learns the reliable weights of each model on an independent validation set. This approach has extremely low computational overhead—the meta-combiner requires only a small number of parameters—and naturally adapts to CPU deployment, while avoiding the difficulty in feature fusion of aligning the CNN's hidden-layer representations with handcrafted cumulant features.

Experimental Results and Statistical Validation

On the RadioML 2016.10a dataset, using 5 random seeds and a held-out test set of 44k frames, FuseAMC achieves an overall accuracy of 55.8%. This is a 5.1 percentage point improvement over the best cumulant-feature baseline and a 3.2 percentage point improvement over the standalone CNN. The lower bounds of the 95% confidence intervals from the paired bootstrap are 5.0 and 3.1 percentage points, respectively, both positive, indicating that the improvements are statistically significant. More importantly, FuseAMC matches the performance of the best classical baseline in the high-SNR regime and matches the CNN's performance in the low-SNR regime—achieving the best of both worlds.

The paired bootstrap is a non-parametric statistical inference method used to assess the significance of the performance difference between two models. It works by resampling the predictions on the test set with replacement; each resampling simultaneously computes the accuracy difference between the two models on the same batch of samples, and after thousands of repetitions, the empirical distribution of the difference is obtained, from which confidence intervals are constructed. Compared with a simple comparison of means, the paired bootstrap accounts for the sampling fluctuations caused by a finite test set and does not rely on normality assumptions. The paper reports that the lower bounds of the 95% confidence intervals are both positive (5.0 and 3.1 percentage points), meaning that in the vast majority of resamplings the fused model outperforms the baselines, and the improvement is not due to random fluctuation—providing statistical evidence that "fusion is effective."

Implications for Edge Deployment

The value of this work lies in the fact that it does not pursue larger and deeper models, but instead achieves significant improvements under CPU-friendly constraints by fusing complementary lightweight methods. For scenarios such as cognitive radio and spectrum monitoring that require real-time processing and are often deployed on embedded platforms, FuseAMC provides a feasible path that does not rely on GPUs. At the same time, it reminds us that classical feature engineering and deep learning are not in a relationship of replacement; when resources are constrained, a carefully designed fusion strategy may be more effective than a single model.

According to the study, FuseAMC strikes a balance among overall accuracy, parameter efficiency, and CPU runnability, offering new ideas for the practical application of automatic modulation classification.

RadioML 2016.10a is a widely used public benchmark dataset in the field of automatic modulation classification, released by DeepSig. It contains 11 modulation types (such as BPSK, QPSK, 8PSK, 16QAM, 64QAM, GFSK, etc.), with SNR ranging from -20dB to +18dB, and each frame consists of 128 complex samples (I/Q two channels). The dataset includes both easily confused dense constellations and higher-order modulations, and has a large proportion of low-SNR samples, so it can effectively test the robustness of models under noisy conditions. It should be noted that this dataset consists of synthetic simulated signals and does not cover multipath, frequency offset, and hardware impairments of real channels. Therefore, the absolute accuracy values obtained on this dataset should not be directly extrapolated to actual deployment scenarios; their value lies more in the relative comparison between methods.

One-line Summary

FuseAMC uses logistic regression to fuse a CNN and a cumulant classifier, achieving AMC performance gains on a single CPU core.

Paragraph Summary

In automatic modulation classification, a compact CNN that directly processes I/Q and a classical classifier based on higher-order cumulants have long lacked systematic comparison. Under controlled conditions, the study finds that the two are complementary in terms of SNR: the CNN is more robust at low SNR, while the cumulant classifier is more accurate at high SNR. Based on this, FuseAMC is proposed, using a logistic regression meta-combiner to fuse the class posterior probabilities of a residual CNN with about 198k parameters and a gradient boosted tree cumulant classifier. The model has about 0.2M parameters and about 3 MFLOPs per frame, and can be trained and inferred on a single CPU core. On RadioML 2016.10a, it achieves 55.8% overall accuracy, 5.1 percentage points higher than the best cumulant baseline and 3.2 percentage points higher than the standalone CNN; the paired bootstrap verifies that the improvement is statistically significant, and it matches the respective best baselines in the high- and low-SNR regimes.