Deep Transfer Learning for Signal Detection in Ambient Backscatter Communications

• Tag detection faces direct-link interference, difficult CSI estimation without ambient-source pilots, and weak backscatter that makes the hypotheses similar despite a BER requirement below 10^-2.Source: §I · pp. 2–4
• The proposed DTL framework transfers offline-learned knowledge to related real-time channel tasks, fine-tuning a DNN to extract channel features and recover tag symbols without explicit channel estimation.Source: §I · p. 4
• The framework derives a DTL-based likelihood ratio test for tag detection according to the minimum error probability criterion.Source: §I · p. 4
• The CMNet detector applies CNNs to sample covariance matrices to extract discriminative features and improve BER performance.Source: §I · p. 4
• The paper derives an asymptotic explicit expression for CMNet properties when the number of samples is sufficiently large.Source: §I · p. 4
• Simulations with modulated and complex Gaussian ambient sources show the proposed algorithm’s BER is close to that of the optimal LRT method with perfect CSI.Source: §I · p. 4

I. INTRODUCTION

• T and H denote transpose and conjugate transpose; CN(µ, Σ) is a circularly symmetric complex Gaussian distribution, while I_M and 0 denote the identity matrix and zero vector.The distribution has mean vector µ and covariance matrix Σ.Source: §I · p. 5
• The notation also defines matrix inverse, max(a, b), the vector 2-norm, and exp(·) as the exponential function.Source: §I · p. 5
• S and T subscripts indicate source and target domains; |·| denotes set cardinality, and Re(·) and Im(·) refer to real and imaginary parts.det(·) denotes the determinant, while P(·) and E(·) denote event probability and statistical expectation.Source: §I · p. 5

II. SYSTEM MODEL

• Each frame contains known tag pilots for training and data symbols for transmission; the channel remains unchanged within a frame.The considered frame structure is designed for slow-fading channels.Source: §II · p. 6
• The reader recovers tag symbols from received observations, which are framed as a binary hypothesis test.The tag can reflect the ambient RF signal or not.Source: §II · p. 7
• The ambient RF source is modeled either as a complex Gaussian signal or as a Q-ary modulated symbol drawn from a constellation.The modulated symbols are drawn with equal probability.Source: §II · p. 7

1) Optimal LRT with Complex Gaussian Ambient Source:

• The likelihood ratio test is considered for a complex Gaussian ambient source.The source signal is modeled as complex Gaussian.Source: 1) Optimal LRT with Complex Gaussian… · p. 8
• The reader must know perfect CSI, including w and h, to use the stated test.The passage identifies w and h as examples of the required CSI.Source: 1) Optimal LRT with Complex Gaussian… · p. 8
• The section gives the logarithmic form of the LRT for the complex Gaussian-source case.The optimal LRT serves as a benchmark for the considered system.Source: 1) Optimal LRT with Complex Gaussian… · p. 8

2) Optimal LRT with Modulated Ambient Source:

• The optimal LRT requires perfect CSI, which may be unavailable in practical AmBC systems; the proposed DNN explores sample-covariance features instead.The covariance matrix implicitly captures distinguishable features such as energies and eigenvalues.Source: 2) Optimal LRT with Modulated Ambient… · pp. 8–9

III. DEEP TRANSFER LEARNING-BASED TAG SIGNAL DETECTION FRAMEWORK

• The DTL framework has offline learning, transfer learning, and online detection stages, with a pre-trained network adapted to current channel coefficients.Partial layers are frozen while remaining layers are fine-tuned before online tag detection.Source: §III · pp. 9–10

A. Initialization for Deep Transfer Learning

• Initialization treats offline and transfer data as distinct source and target domains, using source-domain knowledge to improve target-domain detection.The framework is also valid when the channel coefficients in the two domains come from different distributions, though adaptation takes longer.Source: A. Initialization for Deep Transfer… · pp. 10–11

B. Offline Learning: Pre-trained DNN

• Tag detection is formulated as binary classification, with labels encoding the two hypotheses.The DNN outputs a two-class score vector.Source: B. Offline Learning: Pre-trained DNN · p. 12
• Offline learning maximizes the likelihood, or equivalently the log-likelihood, of the source-domain labels given the features.The objective is expressed as a cost function for estimating the network parameters.Source: B. Offline Learning: Pre-trained DNN · pp. 12–13
• Backpropagation produces the well-trained source-domain DNN used as the pre-trained network for transfer learning.The network's class score represents the likelihood of each category.Source: B. Offline Learning: Pre-trained DNN · pp. 12–13

C. Transfer Learning: Fine-tuning the Pre-trained DNN

• Fine-tuning freezes the transferred layers and updates only the target network's fine-tuning layers using the pilot-based training set.Backpropagation produces the trained target-domain parameters.Source: C. Transfer Learning: Fine-tuning the… · pp. 13–14
• The target network reuses partial feature layers from a pre-trained source-domain DNN, transferring them into the target-domain network.The source and target tasks address the same tag-detection problem.Source: C. Transfer Learning: Fine-tuning the… · p. 13
• Under slow fading, pilots and test data in one tag frame share channel coefficients and statistical distributions, so pilots provide target-domain training data.The authors use this correspondence to train the network on features relevant to test data.Source: C. Transfer Learning: Fine-tuning the… · p. 13
• The trained network's class scores are treated as posterior probabilities, converted to likelihoods with Bayes' theorem, and used in an MEP-based DTL likelihood-ratio test.With equal binary-hypothesis priors, the decision selects H1 when the likelihood ratio exceeds 1; otherwise it selects H0.Source: C. Transfer Learning: Fine-tuning the… · p. 14

D. Online Detection: DTL-LRT using Well-trained DNN

• For online detection, test data is sent to the well-trained DNN, and its DTL likelihood ratio selects H1 above 1 and H0 otherwise.The decision rule is stated for arbitrary test data in the target domain.Source: D. Online Detection: DTL-LRT using… · p. 15
• The DTL workflow is presented as a universal tag-detection scheme that can use different DNN structures and does not explicitly estimate channel coefficients.It requires a few tag-signal pilots for transfer learning.Source: D. Online Detection: DTL-LRT using… · p. 15
• The authors state that the DTL scheme can extend to M-ary tag modulation by changing binary classification to M-class classification.Source: D. Online Detection: DTL-LRT using… · p. 15

IV. CNN-BASED DEEP TRANSFER LEARNING FOR TAG SIGNAL DETECTION

• CMNet uses a sample covariance matrix as input because it captures distinguishable features, which CNNs can extract from matrix-formed data.The complex-valued input is represented using real and imaginary channels.Source: §IV · p. 15, A. CMNet Structure · p. 17

A. CMNet Structure

• CMNet combines convolutional and pooling feature extraction with dropout and fully connected layers for nonlinear feature combinations.Dropout is included because transfer learning has limited training examples and may otherwise overfit.Source: A. CMNet Structure · pp. 16, 18
• The authors state that CMNet can be adapted to different sample-covariance-matrix shapes as antenna counts change.Source: A. CMNet Structure · p. 18

1) Offline Learning:

• For each source-domain tag symbol, the method forms a sample covariance matrix from its sampling matrix and pairs it with the corresponding binary label.Source: 1) Offline Learning · p. 19
• The source-domain training set is used to train CMNet with a cost function and backpropagation, yielding a pre-trained network.Source: 1) Offline Learning · pp. 19–20
• Offline learning uses well-developed channel models to obtain training data for the source domain.The authors describe these models as characterizing actual channels in terms of channel statistics.Source: 1) Offline Learning · p. 19

2) Transfer Learning:

• Target-domain training data is formed from pilot samples by constructing sample covariance matrices and pairing them with tag-symbol labels.Source: 2) Transfer Learning · p. 20
• Transfer learning freezes the pre-trained convolutional layers and updates only the fully connected layers using backpropagation.The target-domain cost function guides this fine-tuning.Source: 2) Transfer Learning · p. 20

3) Online Detection:

• For online detection, the trained CMNet produces a likelihood-ratio test on the test sample covariance matrix, with a threshold of 1 determining the binary decision.Source: 3) Online Detection · p. 20

4) Algorithm Steps:

• The paper summarizes the CMNet-based detection procedure in an algorithm, with separate iteration indices and maximum iteration counts for source and target training.Source: 4) Algorithm Steps · p. 21

C. Theoretical Analysis

• For sufficiently large N in a richly scattered multipath environment, CMNet’s output-element ratio matches the perfect-CSI optimal LRT expression.The expression includes the detection threshold γ.Source: C. Theoretical Analysis · pp. 22, 24
• Under rich scattering, the perfect-CSI optimal LRT behaves as an energy detector.Source: C. Theoretical Analysis · p. 22
• As N grows, the sample covariance matrix approaches a real-valued diagonal matrix, while its imaginary part approaches zero.Source: C. Theoretical Analysis · p. 22
• The stated channel assumption is for analytical convenience; the proposed method is valid for channel models meeting the previously stated conditions.Source: C. Theoretical Analysis · p. 22

V. NUMERICAL RESULTS

• CMNet’s online detection time is 2.6 milliseconds on the specified CPU-and-GPU desktop, shorter than the coherence time of a general slow-fading environment.Source: §V · p. 24
• At BER ≈10^-2, CMNet gains 4 dB SNR over traditional SVM, outperforms ED and SVM, and nearly matches the optimal LRT.The ED comparator uses perfect CSI.Source: §V · p. 25, Figure 7 · p. 25
• CMNet achieves BER close to the optimal LRT under both QPSK and complex Gaussian ambient sources, although all tested algorithms perform slightly worse for the Gaussian source.Source: §V · p. 26, Figure 8 · p. 26
• As N increases, CMNet’s BER scales with the optimal LRT’s slope, and its performance remains scalable across antenna counts, with gains at larger M.For sufficiently large M, the method exploits spatial degrees of freedom for detection.Source: §V · pp. 26–27
• Increasing reflection coefficient ζ improves BER for each algorithm, while CMNet remains near the optimal LRT and outperforms ED and SVM.The paper attributes the improvement to a stronger reflected path that makes tag signals easier to distinguish from direct-path signals.Source: §V · p. 27
• At a BER requirement of 10^-2, the proposed method supports a tag-to-reader distance of 2 m and approaches optimal-LRT performance.Source: Figure 9 · p. 27, §V · p. 28

VI. CONCLUSIONS

• The DTL framework uses a pretrained DNN and a few pilots for transfer learning, producing a DTL-LRT detector that recovers tag symbols without explicit channel estimation.Source: §VI · p. 28
• The paper applies CNN feature extraction to the sample covariance matrix in a CMNet detector and provides theoretical analysis of the method.Source: §VI · p. 28
• CMNet achieves close-to-optimal performance without explicitly obtaining CSI for modulated and complex Gaussian ambient signals.Source: §VI · p. 28