---
title: 'RIS-MAE: Self-Supervised AMC via Masked Autoencoder'
url: https://www.emergentmind.com/topics/ris-mae
type: topic
---

# RIS-MAE: Self-Supervised AMC via Masked Autoencoder

RIS-MAE is a self-supervised automatic modulation classification (AMC) framework designed to learn robust modulation representations directly from raw IQ sequences using a masked autoencoder. It is intended to address two recurrent limitations in deep AMC: over-reliance on time-frequency images and dependence on large labeled datasets. The method performs self-supervised pre-training on unlabeled raw IQ data and then fine-tunes the learned encoder on a small labeled subset. In the reported evaluation on four datasets, it performs better than existing methods in few-shot and cross-domain tasks, with particular emphasis on transfer to previously unseen datasets under limited fine-tuning data [2508.00274].

## 1. Nomenclature and scope

In the title “RIS-MAE: A Self-Supervised Modulation Classification Method Based on Raw IQ Signals and Masked Autoencoder,” the term denotes a raw-signal masked-autoencoding framework for AMC [2508.00274]. That explicit usage is distinct from several unrelated uses of similar shorthand in adjacent reconfigurable-intelligent-surface literature.

The acronym environment is not uniform. “MA-RIS” denotes a movable-element RIS in a downlink RIS-assisted link [2405.00141]. A summary of “Movable Antenna-enabled RIS-aided Integrated Sensing and Communication” uses “RIS-MAE” for a movable antenna-enabled RIS-aided ISAC system [2407.03228]. A summary of “Opportunistic Subarray Grouping for RIS-Aided Massive Random Access in Cellular Connectivity” uses “RIS-MAE” for RIS-aided massive random access [2503.17904]. A summary of “RIS-assisted Atomic MIMO Receiver” states that the paper itself primarily uses the title “RIS-assisted Atomic MIMO Receiver” and does not explicitly introduce a separate acronym “RIS-MAE” beyond that framing [2510.15763].

Accordingly, the clearest explicit definition of RIS-MAE is the modulation-classification framework of [2508.00274]. The remaining usages belong to different research lines centered on reconfigurable intelligent surfaces rather than on masked autoencoding for raw IQ sequences.

## 2. Signal-domain formulation

RIS-MAE is built on the premise that a received wireless signal is naturally represented as a complex-valued IQ sequence,
\[
x_i = I_i + jQ_i,
\]
where \(I_i\) and \(Q_i\) are the in-phase and quadrature components [2508.00274]. The AMC task is written as
\[
f: x \rightarrow y,\quad y \in M,
\]
where \(x\) is the received IQ sequence, \(y\) is the modulation label, and \(M\) is the set of modulation classes. The signal model is
\[
s(t) = F(x(t)) \cdot e^{j(2\pi \Delta f t + \phi)},
\]
with \(F(\cdot)\) the modulation function, \(\Delta f\) the carrier frequency offset, and \(\phi\) the phase offset; after channel effects,
\[
r(t) = h(t) * s(t) + n(t),
\]
where \(h(t)\) is the channel impulse response, \(*\) is convolution, and \(n(t)\) is AWGN noise.

The framework argues that raw IQ signals preserve key modulation cues that are often degraded by intermediate transforms. The paper explicitly identifies amplitude, phase, temporal dependencies, frequency offset, multipath/fading structure, and symbol transitions as signal-domain properties retained in raw IQ but liable to degradation when AMC pipelines first convert signals into spectrogram-like inputs using FFT, STFT, or CWT [2508.00274]. The stated concern is that such conversions can discard phase continuity, frequency offset, symbol structure, and multipath effects, thereby reducing adaptability to different communication conditions.

This raw-signal orientation is central to RIS-MAE’s transfer claim. Because the model is trained directly on the native signal representation rather than on a derived image representation, it is intended to learn features that remain useful under few-shot learning, cross-domain generalization, real-world deployment, and low-SNR robustness [2508.00274].

## 3. Architecture and learning pipeline

RIS-MAE follows a two-stage pipeline: self-supervised pre-training and supervised fine-tuning [2508.00274]. In pre-training, the input is unlabeled raw IQ sequences, and the task is to randomly mask a large fraction of patches and reconstruct them. In fine-tuning, the pre-trained encoder is reused with a simple linear classifier on top.

The paper uses raw IQ signals of length \(1024\), represented as
\[
X = \begin{bmatrix}
I_1 & I_2 & \cdots & I_{1024} \\
Q_1 & Q_2 & \cdots & Q_{1024}
\end{bmatrix} \in \mathbb{R}^{2 \times 1024}.
\]
The sequence is split into \(N = 128\) patches with patch size \(8\), and each patch contains both I and Q channels. The \(i\)-th patch is projected into a latent vector by
\[
Z_i = W_{\text{patch}} \cdot X_{\text{patch},i} + b_{\text{patch}}, \quad i=1,2,\dots,N,
\]
with \(W_{\text{patch}} \in \mathbb{R}^{d \times 16}\) and \(d = 768\), followed by a learnable positional embedding,
\[
Z_i = Z_i + p_i,\quad p_i \in \mathbb{R}^d.
\]
The use of 1D patches, rather than image-style 2D patches, is motivated by the fact that IQ signals are time series and 1D patchification better preserves temporal continuity [2508.00274].

During pre-training, RIS-MAE applies random masking to patches. If \(p\) is the masking ratio, then
\[
|M| = pN,
\]
and the visible patches are
\[
Z_{\text{keep}} = \{z_i \mid i \notin M\}, \quad |Z_{\text{keep}}| = (1-p)N.
\]
A class token is appended,
\[
Z^{(0)} = [Z_{\text{CLS}}; Z_{\text{keep}}] \in \mathbb{R}^{((1-p)N+1)\times d}.
\]
The reported best masking ratio is \(75\%\), which the paper describes as the best balance between enough missing content to force meaningful representation learning and enough visible context to reconstruct successfully [2508.00274].

The encoder is a 12-layer Transformer with LayerNorm, Multi-Head Attention, MLP, and residual connections. The decoder is a 4-layer Transformer decoder with \(d_{\text{dec}} = 512\). After reinserting mask tokens and restoring the original token order, the decoder predicts the masked patches. The pre-training loss is the masked-patch reconstruction MSE,
\[
L_{\text{MAE}} = \frac{1}{|M|} \sum_{i\in M} \|\hat{x}_i - x_i\|^2,
\]
and the fine-tuning loss is cross-entropy,
\[
L_{\text{CE}} = -\sum_{c=1}^{C} y_c \log(p_c).
\]
For downstream classification, the class token representation is
\[
h_{\text{CLS}} = Z_{\text{enc,CLS}},
\]
and the classifier output is
\[
y = \mathrm{softmax}(W_{\text{cls}} \cdot h_{\text{CLS}} + b_{\text{cls}}).
\]

The experimental setup reported for training is specific. Pre-training uses only samples with SNR \(\ge 6\) dB, AdamW, learning rate \(1\times 10^{-4}\), warm-up \(5\%\), 100 epochs, and batch size 1024. Fine-tuning uses all SNR ranges, AdamW, learning rate \(1\times 10^{-4}\), warm-up \(10\%\), 50 epochs, and batch size 512. The data split is \(80\%\) for self-supervised pre-training and \(20\%\) for supervised fine-tuning; for pre-training datasets, \(70\%\) is train and \(30\%\) validation, while fine-tuning uses \(50\%\) validation and \(50\%\) test [2508.00274].

## 4. Evaluation protocol and empirical results

RIS-MAE is evaluated on four datasets: RML2018.01a, HisarMod2019.1, CommData, and Panoradio [2508.00274]. RML2018.01a contains 2,555,904 samples, 24 modulation types, SNR from \(-20\) dB to \(+30\) dB in 2 dB steps, and 1024 IQ samples each. HisarMod2019.1 contains 780,000 samples, 26 modulation schemes, SNR from \(-20\) dB to \(+18\) dB, and includes five channels: ideal, static, Rayleigh fading, Rician fading (\(k=3\)), and Nakagami-m fading (\(m=2\)). CommData contains 27,470,000 samples, 14 modulation schemes, SNR from \(-20\) dB to \(+20\) dB, 1024 samples each, is collected via USRP hardware, and includes frequency offset, multiple symbol rates, and multiple bandwidths. Panoradio contains 172,800 samples, 18 modulation schemes, SNR from \(-10\) dB to \(+25\) dB, 2048 samples each downsampled to 1024 in experiments, and includes unseen modulation types such as MT63_1000, USB, LSB, and radiofax, with fading via the Watterson channel model.

The baselines are six supervised methods: CNN2, IC-AMCNet, MCLDNN, MCNET, PET-CGDNN, and ResNet. The evaluation metrics are Overall Accuracy (OA) and Kappa coefficient. In the key few-shot setting of \(1\%\) labeled data for fine-tuning, RIS-MAE reports OA values of 48.41 on RML2018.01a, 40.26 on HisarMod2019.1, 34.31 on CommData, and 46.38 on Panoradio. The corresponding Kappa values are 0.4616, 0.3787, 0.2913, and 0.4322. The reported OA improvements over baselines are at least 11.81% on RML2018.01a, at least 5% on HisarMod2019.1, at least 6.93% on CommData, and at least 8.09% on Panoradio [2508.00274].

Cross-domain generalization is emphasized through Panoradio, which was not used during pre-training. After fine-tuning with very few labels, the paper reports strong recognition for unseen classes: MT63_1000 at 96%, USB at 92%, LSB at 97%, and FAX at 68% [2508.00274]. The ablation study identifies the best masking ratio as \(75\%\), the best pre-training SNR threshold as SNR \(\ge 6\) dB, and the best patch size as 8. It also states that RIS-MAE is especially strong when labeled data is extremely scarce, specifically at 1%, 0.5%, and 0.1%, and that the advantage increases as labels become scarcer [2508.00274].

## 5. Failure modes, limitations, and interpretive significance

The paper identifies several limitations. High-order modulations remain difficult; confusions are common among high-order QAMs and high-order PSKs because constellation points become closer together and more noise-sensitive [2508.00274]. Similar modulation families are also hard to separate, with specific examples including 8ASK vs OQPSK, AM-SSB-WC vs 64QAM, and PSK31 vs PSK63. On CommData, performance is described as slightly lower because of hardware noise, nonlinear distortion, clock drift, and symbol-rate and bandwidth variation.

These limitations qualify the scope of the reported transfer benefits. RIS-MAE does not eliminate class overlap or hardware-induced variability, but the reported few-shot and cross-domain results indicate that self-supervised pre-training on raw IQ sequences can improve representation quality under constrained labeling regimes [2508.00274]. This suggests that the principal strength of the method lies in representation transfer and label efficiency rather than in complete resolution of intrinsically similar waveform families.

The practical motivation stated for the framework is equally specific. The method is positioned for spectrum monitoring, cognitive radio, and secure communications, with the broader claim that self-supervised pre-training on raw IQ sequences is promising for real-world deployment because it avoids expensive time-frequency preprocessing, reduces dependence on labeled data, and learns physically meaningful signal representations [2508.00274].

## 6. Relation to reconfigurable-intelligent-surface research

A recurrent misconception arises from the string “RIS.” In RIS-MAE [2508.00274], the term names a masked-autoencoding AMC framework and is not a study of electromagnetic reconfigurable intelligent surfaces. By contrast, “RIS-aided Wireless Communications: Can RIS Beat Metal Plate?” studies whether a well-configured RIS can be replaced by a well-rotated metal plate of the same size and shape, and concludes that a well-rotated metal plate can only approximately match RIS performance under limited conditions; when the incident field is spherical and the links are not very far away, a well-configured RIS outperforms a metal plate of the same physical size [2303.02938].

Other nearby papers similarly use RIS in the electromagnetic sense. “RIS-aided Wireless Communication with Movable Elements Geometry Impact on Performance” studies MA-RIS and reports 24% improvement in outage probability and 2 dB gain in Signal-to-Noise Ratio relative to FPA-RIS [2405.00141]. “Movable Antenna-enabled RIS-aided Integrated Sensing and Communication” investigates a movable antenna-aided ISAC system with a joint optimization over BS beamforming, RIS reflecting coefficients, and antenna positions, and reports that the MA and RIS-aided ISAC system outperforms conventional fixed position antenna and RIS-aided systems [2407.03228]. “Opportunistic Subarray Grouping for RIS-Aided Massive Random Access in Cellular Connectivity” formulates a MAC-layer sequential decision problem for RIS probing, subarray grouping, and data transmission, and reports that the proposed strategy significantly outperforms the extremes of full-array grouping and element-wise grouping [2503.17904]. “RIS-assisted Atomic MIMO Receiver” uses RIS and PAM to align the phase of the transmitted signal with that of the local oscillator, thereby mitigating phase ambiguity and substantially reducing both signal detection complexity and overall receiver complexity [2510.15763].

The distinction is therefore categorical rather than cosmetic. RIS-MAE [2508.00274] belongs to self-supervised learning for modulation classification on raw IQ sequences, whereas the other RIS-centered works concern wave scattering, geometry control, MAC-layer access design, sensing, or receiver architecture. The shared acronymal surface conceals substantially different objects of study.

Source: https://www.emergentmind.com/topics/ris-mae