---
title: Centralized Logarithmic Power Spectrum (CLPS)
url: https://www.emergentmind.com/topics/centralized-logarithmic-power-spectrum-clps
type: topic
---

# Centralized Logarithmic Power Spectrum (CLPS)

Searching arXiv for CLPS and related logarithmic power-spectrum work.
Search query: "Centralized Logarithmic Power Spectrum CLPS arXiv"
Centralized Logarithmic Power Spectrum (CLPS) is a channel-resistant spectral fingerprint used in radio frequency fingerprinting (RFF) to identify transmitters from the received preamble while suppressing dominant channel-dependent effects. In the formulation studied in “Collusion-Driven Impersonation Attack on Channel-Resistant RF Fingerprinting,” CLPS is obtained from the received signal by estimating a power spectrum from the FFT of the autocorrelation, taking a logarithm, and then subtracting the mean across frequency bins. The intended effect is to preserve transmitter-specific hardware distortions while removing a channel-dependent bias term; precisely this transferability across channels is also the basis of a cross-channel impersonation attack [2509.22154].

## 1. Definition and formal construction

CLPS is introduced in the context of channel-resistant RFF. The received signal is modeled as
\[
r(t) = f(s(t)) * h(t) + n(t)
\]
where \(s(t)\) is the transmitted baseband signal, \(f(\cdot)\) models transmitter hardware imperfections, \(h(t)\) is the wireless channel impulse response, \(n(t)\) is additive white Gaussian noise, and \(*\) denotes convolution. The RFF task is to extract features that reflect \(f(s(t))\) while being robust to \(h(t)\) and \(n(t)\) [2509.22154].

CLPS is presented as a channel-resistant spectral fingerprint derived from the received preamble. It is intended to preserve spectral signatures introduced by transmitter hardware impairments while removing the dominant channel-dependent spectral offset. The construction combines three operations. The power spectrum captures frequency-domain energy structure. The logarithm compresses dynamic range and converts multiplicative effects into additive ones. Centralization subtracts the mean across frequency bins. In the paper’s terminology, this yields a “centralized logarithmic power spectrum.”

If \(P_r(\omega)\) denotes the power spectrum of \(r(t)\), the logarithmic power spectrum is written as
\[
P_{r}(\omega)_{\log} = \log(\frac{1}{T}) + 2*\log|\text{FFT}(f(s(t)))|  + 2*\log|\text{FFT}(h(t))|
\]
where \(T\) is the duration of the captured signal. Under the assumption that over a short preamble interval the channel is approximately constant, the mean-subtracted logarithmic power spectrum is
\[
\text{CLPS}(\omega) = P_{r}(\omega)_{\log} - \text{mean}(P_{r}(\omega)_{\log})
\]
and then simplifies to
\[
\text{CLPS}(\omega) =2*(\log|\text{FFT}(f(s(t))|  -mean(\log|\text{FFT}(f(s(t))|))
\]
despite typographical inconsistencies in the parentheses. The stated intended meaning is that centralization removes the channel-dependent term and leaves the hardware-dependent spectral structure up to mean removal [2509.22154].

The algorithmic extraction procedure is given explicitly. First, compute the autocorrelation of the received signal \(\mathbf{R}\):
\[
\mathbf{ACF} \gets \text{Autocorrelation}(\mathbf{R})
\]
Second, compute the power spectrum:
\[
\mathbf{P}_{\text{spec}} \gets |\mathcal{FFT}(\mathbf{ACF})|^2
\]
Third, take the logarithm:
\[
\mathbf{P}_{\log} \gets \log_{10}(\mathbf{P}_{\text{spec}})
\]
Fourth, centralize:
\[
\mathbf{CLPS} \gets \mathbf{P}_{\log} - \text{Mean}(\mathbf{P}_{\log})
\]
Operationally, CLPS is therefore a mean-centered log-power spectrum of the autocorrelation spectrum of the received preamble.

## 2. Channel resistance and the underlying system model

The central rationale for CLPS is spectral factorization under convolution. Because the channel appears multiplicatively in the frequency domain, taking log-magnitude converts multiplicative factors into additive ones. If the preamble is short enough that the channel is approximately constant over the captured interval, then subtracting the mean can reduce or remove the dominant channel contribution. The paper states this explicitly: “This formulation removes the channel-dependent term and preserves the spectral features introduced by hardware imperfections” [2509.22154].

Within this framework, CLPS is intended to be robust to additive white Gaussian noise, multipath fading, and Doppler shift, provided the preamble is fixed and channel variations over the preamble are limited enough for the approximation to hold. The simulation model retains the same received-signal form,
\[
r(t)=f(s(t))*h(t)+n(t),
\]
and treats AWGN, Rician fading, Rayleigh fading as the \(K=0\) special case of Rician, multipath fading with six taps, and Doppler shift as channel effects contained in \(h(t)\) or \(n(t)\) [2509.22154].

The default training channel parameters are stated as Rician + AWGN, with delay vector \([0, 50, 110, 170, 290, 310]\) ns, gain vector \([0, -3, -10, -18, -26, -32]\) dB, \(K\)-factor \(5\), maximum Doppler shift \(10\) Hz, and \(E_b/N_0 = 10\) dB. Evaluated environments include AWGN, Indoor Office Channel A, Indoor Office Channel B, and Vehicular Channel A, using ITU-R M.1225-style models. The article’s technical interpretation is careful: CLPS is treated as channel-resistant rather than channel-immune [2509.22154].

## 3. CLPS in the collusion-driven impersonation threat model

The exact term CLPS is studied most directly in a security setting. The threat model contains four roles: legitimate transmitter(s), a victim receiver, an attacker, and a colluding receiver. Legitimate transmitters periodically send packets with a fixed preamble and exhibit unique hardware distortions, producing distinctive CLPS fingerprints. The victim receiver extracts CLPS from the preamble and identifies the transmitter using a deep neural network classifier. The attacker seeks to impersonate a specific transmitter without knowing the exact channel to the victim and without fully knowing how its own hardware will distort the crafted signal after transmission. The colluding receiver is located elsewhere, receives both target and attacker transmissions, and runs the same CLPS extraction and the same classifier as the victim receiver [2509.22154].

The key attack insight is that if the attacker can generate transmissions whose CLPS matches that of the legitimate transmitter at the colluding receiver, then the victim receiver is also likely to accept the forged fingerprint. Earlier collusion attacks under AWGN relied on the assumption that failure of discrimination at the colluder implied failure at the victim. The paper argues that this assumption breaks under multipath when the feature is not channel-resistant. CLPS changes the situation because it is intended to produce approximately the same transmitter fingerprint under different channels. The colluder therefore acts as a surrogate observer of the victim’s feature space not because the channels are identical, but because CLPS is designed to remove channel dependence [2509.22154].

This yields a structural conclusion about CLPS. Its strength for identification is also its weakness under collusion. A feature engineered to be stable across locations can become easier to mimic across locations when the attacker has feedback from a colluding receiver. The paper states the practical implication directly: channel robustness does not imply security.

## 4. Generative spoofing methodology centered on CLPS matching

The attack is implemented with a VAE-based spoofed signal generation network. The attacker uses perturbed baseband preambles as input, and the network outputs reconstructed or spoofed waveform-domain signal segments. The relevant objects are \(r(t)_{input}\), \(r(t)_{recon}\), \(CLPS_{recon}\), \(CLPS_{target}\), \(y_{recon}\), and \(y_{target}\) [2509.22154].

The VAE architecture contains an encoder with 3 convolutional layers, kernel size \((1 \times 5)\), stride 2, and ReLU activations; a latent representation with two fully connected layers outputting mean and log-variance; latent dimension \(128\); a sampling module that samples a latent variable from the learned Gaussian; and a decoder with one fully connected layer and three transposed convolutional layers that reconstruct the waveform to original shape. The latent variables are the standard \(\mu\) and \(\sigma^2\), with a KL term that pushes the approximate posterior toward a standard normal prior [2509.22154].

The full loss is
\[
Loss = {\lambda}_{recon} * loss_{recon} + {\lambda}_{KL} *loss_{KL} + {\lambda}_{CLPS}*loss_{CLPS}+{\lambda}_{Cls}*loss_{Cls}
\]
with \(\lambda_{recon}=2.0\), \(\lambda_{KL}=0.1\), \(\lambda_{CLPS}=1.0\), and \(\lambda_{Cls}=0.5\). The component terms are
\[
loss_{recon} = ||r(t)_{recon} - r(t)_{input}||_2^2
\]
\[
loss_{KL} = -0.5\sum(1 + \log(\sigma^2) - \mu^2 - \sigma^2)
\]
\[
loss_{CLPS} = ||CLPS_{recon} - CLPS_{target}||_2^2
\]
\[
loss_{Cls} = CrossEntropy(y_{recon},y_{target})
\]
The CLPS loss is the central feature-matching term, and the classification loss is the decision-level deception objective. The paper’s motivation for using a VAE rather than a GAN is that the distribution of CLPS features is relatively concentrated and has a clear underlying structure; in their account, this makes GAN training vulnerable to discriminator saturation and vanishing gradients, whereas the VAE provides explicit losses and stable gradients [2509.22154].

## 5. Experimental setting, empirical behavior, and limitations

The reported experiments use 10 legitimate transmitters and 1 attacker. Device-specific impairment is modeled by I/Q imbalance, with gain imbalance in \([-0.3, 0.3]\) and phase imbalance in \([-15^\circ, 15^\circ]\). The signal uses an IEEE 802.15.4 fixed preamble with sample length 5120 complex I/Q samples. The dataset contains 55,000 training samples and 5,500 test samples; each legitimate device contributes 5000 training segments and 500 testing segments. CLPS is extracted from each segment, and identification uses a ResNet-based classifier following the architecture in the referenced CLPS work [2509.22154].

Three baselines are compared: I/Q-domain impersonation without CLPS, GAN-based spoofed signal generation, and the proposed CLPS + VAE + multi-objective loss. The evaluation metric is Attack Success Rate (ASR), defined as the proportion of attacker transmissions that are misidentified as the target legitimate device. The empirical pattern is consistent throughout the paper. Replacing CLPS with raw I/Q features fails badly because I/Q is highly channel-sensitive and the colluder cannot provide stable feedback under multipath. GAN-based generation in CLPS space also fails, with spoofed samples consistently misclassified as the wrong device rather than the intended target. The VAE-based method succeeds because it directly optimizes CLPS alignment and classifier deception [2509.22154].

The headline quantitative claim is that the proposed method essentially maintains a success rate above \(95\%\) under different channel conditions. The paper also gives explicit figures. Under varying \(K\)-factor in Indoor Office Channel A, ASR is approximately \(97.40\%\) at \(K=0\) and rises to nearly \(100.0\%\) at \(K=10\). In Indoor Office Channel B, ASR is \(0.20\%\) at \(K=0\) and rises to \(100.0\%\) as conditions improve. In Vehicular Channel A, ASR is \(12.40\%\) at \(K=0\) and rises to \(99.80\%\) at \(K=10\). Across devices in cross-channel tests, most target devices under non-AWGN-trained models are often spoofed with ASR nearing \(100\%\), while some mismatched settings drop to around \(80\%\). As \(E_b/N_0\) increases, ASR increases; for \(E_b/N_0 \ge 0\), the attack shows good robustness, and the vehicular channel performs worse at low \(E_b/N_0\) before quickly approaching \(100\%\) [2509.22154].

The same results establish a common misconception to avoid: CLPS is not presented as perfectly invariant. The paper states that the attack degrades when the colluder and victim experience more divergent channels, when low \(K\)-factor creates strong NLoS multipath, and when low \(E_b/N_0\) increases noise distortion. The correct characterization is therefore channel-resistant, not channel-immune.

## 6. Broader context and related logarithmic power-spectrum constructions

Outside RF fingerprinting, several arXiv papers study logarithmic transformed power spectra that are conceptually related but do not use the exact term “Centralized Logarithmic Power Spectrum.” In cosmology, “Rejuvenating the matter power spectrum: restoring information with a logarithmic density mapping” defines \(A=\log(1+\delta)\), explicitly notes that \(A\) is not a zero-mean field, and uses \(\log(1+\delta)-\langle \log(1+\delta)\rangle\) in covariance estimation [0903.4693]. “Gaussianization: Enhancing the Statistical Power of the Power Spectrum” studies the power spectrum of \(\ln(1+\delta)\) and argues that a log transform can dramatically reduce power-spectrum covariance for the roughly lognormal low-redshift matter field [1109.3476]. “Log-transforming the matter power spectrum” develops exact relations between a log-normal field and the power spectrum of its logarithmic Gaussian field, while treating the mean or monopole as a separate issue that may be fixed by setting the monopole to zero [1312.1354]. For discrete galaxy surveys, “The Bias of the Log Power Spectrum for Discrete Surveys” uses the centralized variables \((A-\overline A)\) and \((A^*-\overline{A^*})\) in its two-point statistics and derives a large-scale multiplicative bias between \(P_{A^*}(k)\) and \(P_A(k)\) [1708.00954].

These cosmological constructions are not the same object as RF CLPS. In the RF setting, centralization removes a channel-dependent bias term across frequency bins of a received preamble-derived spectrum. In the cosmological setting, the closely related operation is mean subtraction of a log-transformed field before Fourier analysis. The similarity lies in the shared logic of taking logarithms to reshape multiplicative structure and using centering to control a dominant offset. A plausible implication is that the term CLPS names a specific RF fingerprint feature, whereas “CLPS-like” transformed spectra in other fields denote analogous but formally distinct mean-subtracted logarithmic power-spectrum constructions.

Within RF security, the most important conclusion remains the one established by the CLPS attack paper. CLPS is a strong channel-resistant RFF feature for identification, but its structured, concentrated, and transferable nature makes it vulnerable to collusion-driven feature-space mimicry. The defenses explicitly suggested are adversarial training of the classifier against synthesized spoofed samples and active defense at the legitimate transmitter, such as embedding recoverable perturbations known only to the intended receiver [2509.22154].

Source: https://www.emergentmind.com/topics/centralized-logarithmic-power-spectrum-clps