Encrypted Semantic Super-Resolution Communication

• The paper introduces a secure semantic communication framework that jointly integrates semantic extraction, deep JSCC, and encryption to protect image features during transmission.
• It employs a lightweight modulo-256 encryption method on semantic features, ensuring that intercepted data remains unintelligible without a pre-shared key under AWGN conditions.
• Receiver-side super-resolution using an RDN module significantly boosts PSNR in low-SNR environments by compensating for errors introduced through compression and encryption.

Searching arXiv for the specified paper and closely related context papers to ground the article. Encrypted Semantic Super-Resolution Enhanced Communication (SREC) is a secure semantic communication framework for end-to-end image transmission that combines semantic extraction and deep joint source-channel coding (JSCC), lightweight feature encryption, and receiver-side super-resolution reconstruction. Its stated objective is to protect semantic features from eavesdropping while preserving reconstruction quality under noisy wireless conditions, particularly in an additive Gaussian white noise (AWGN) channel. In SREC, an input image is converted into semantic features, mapped to channel symbols, encrypted by a modulo-256 operation, transmitted, decrypted at the receiver, reconstructed into an image, and then refined by a Residual Dense Network (RDN)-based super-resolution module (Zhang et al., 5 Sep 2025).

1. Conceptual basis and system scope

SREC is situated within semantic communication (SemCom), which is described as a paradigm of deep integration between artificial intelligence and communication technology. In this setting, the transmitted object is not a pixel-faithful source bitstream but a learned semantic representation optimized for downstream reconstruction quality. The motivating problem is that semantic features sent in plaintext over physical channels are vulnerable to interception. SREC addresses that vulnerability by inserting encryption directly into the semantic communication pipeline rather than treating security as a separate post hoc layer (Zhang et al., 5 Sep 2025).

The framework combines three elements in one image-transmission system:

1. semantic extraction and JSCC transmission,
2. lightweight feature encryption,
3. receiver-side super-resolution reconstruction.

This combination is central to the formulation. The semantic encoder and JSCC modules compress and adapt the image representation to the channel; encryption protects the latent representation during transmission; and super-resolution compensates for quality degradation induced by channel noise, encryption/decryption-related errors, and the reduced fidelity of compressed semantic features.

The paper explicitly targets image transmission. Its empirical claims are framed around PSNR, AWGN robustness, different digital modulation schemes, and eavesdropping resistance under a shared-key assumption. It therefore belongs simultaneously to secure semantic communication, learned JSCC, and learning-based image reconstruction (Zhang et al., 5 Sep 2025).

2. End-to-end architecture and signal flow

The transmitter–channel–receiver chain is defined as a sequence of semantic, communication, cryptographic, and reconstruction stages. The main functional mappings are

$y = G_a(x),$

$s = F_e(y),$

$s_{\text{enc}} = Enc(s),$

$s_{\text{trans}} = h * s_{\text{enc}} + n,$

$s_{\text{dec}} = Dec(s_{\text{trans}}),$

$\hat{y} = F_d(s_{\text{dec}}),$

$\hat{x} = G_s(\hat{y}),$

$\hat{x}_{\text{sr}} = RDN(\hat{x}),$

where $G_a(\cdot)$ is semantic extraction, $F_e(\cdot)$ is the JSCC encoder, $s = F_e(y),$0 is the encryption function, $s = F_e(y),$1 is the channel coefficient, $s = F_e(y),$2 is additive Gaussian noise, $s = F_e(y),$3 is decryption, $s = F_e(y),$4 is the JSCC decoder, $s = F_e(y),$5 is semantic recovery, and $s = F_e(y),$6 is the super-resolution network (Zhang et al., 5 Sep 2025).

Operationally, the algorithmic pipeline is described as:

1. input image $s = F_e(y),$7,
2. semantic extraction to obtain $s = F_e(y),$8,
3. JSCC encoding to produce $s = F_e(y),$9,
4. encryption with a pre-shared key to obtain $s_{\text{enc}} = Enc(s),$0,
5. transmission over the wireless channel,
6. decryption to obtain $s_{\text{enc}} = Enc(s),$1,
7. JSCC decoding to estimate $s_{\text{enc}} = Enc(s),$2,
8. semantic recovery to reconstruct $s_{\text{enc}} = Enc(s),$3,
9. super-resolution enhancement to produce $s_{\text{enc}} = Enc(s),$4.

A central design point is that the system does not transmit raw pixels. Instead, it transmits semantic features encoded into channel symbols by a deep JSCC model. In the reported experiments, SREC adopts NTSCC as the semantic extraction and JSCC backbone. The paper characterizes this as consistent with JSCC-style semantic communication, where transmission is optimized for reconstruction quality rather than exact bitwise source recovery.

3. Encryption mechanism, decryption, and security interpretation

The encryption stage is intentionally lightweight. After JSCC encoding, the normalized feature tensor $s_{\text{enc}} = Enc(s),$5 is encrypted element-wise with a pseudorandom key tensor of the same size. The key is defined as

$s_{\text{enc}} = Enc(s),$6

The encryption rule is

$s_{\text{enc}} = Enc(s),$7

The paper emphasizes that the key is a pseudorandom tensor of type torch.uint8, that legitimate parties share it in advance, and that the key is transmitted through a secure channel, while the encrypted features traverse the public wireless channel (Zhang et al., 5 Sep 2025).

At the receiver, decryption is the modular subtraction counterpart of encryption:

$s_{\text{enc}} = Enc(s),$8

The security rationale is that semantic features are never exposed in plaintext during transmission. Even if an eavesdropper intercepts $s_{\text{enc}} = Enc(s),$9, the paper argues that the latent representation cannot be converted into meaningful image semantics without the pre-shared key. The threat model further assumes that an adversary may know all model parameters but does not know the key.

The security claim is explicitly limited. The paper states semantic security intuition based on the difficulty of distinguishing pseudorandom sequences from random sequences in polynomial time, and it references scalable encrypted image coding work for that interpretation. At the same time, it does not provide a formal cryptographic proof beyond that intuition. A common misconception would be to equate the modulo-256 feature masking step with a complete cryptographic analysis; the paper instead presents it as a lightweight, JSCC-compatible protection mechanism under a secure key-distribution assumption (Zhang et al., 5 Sep 2025).

4. Receiver-side super-resolution and the RDN module

The super-resolution component is introduced to mitigate three degradation sources identified by the paper: wireless channel noise, error propagation through encryption and decryption, and the reduced fidelity of compressed semantic features. It is placed after semantic recovery, taking the reconstructed image $s_{\text{trans}} = h * s_{\text{enc}} + n,$0 and producing the refined output $s_{\text{trans}} = h * s_{\text{enc}} + n,$1.

The stated intuition is that JSCC decoding and semantic recovery may preserve coarse structure while losing fine detail, especially under low SNR. The super-resolution network supplies a learned image prior that reconstructs plausible high-frequency content and sharpens textures. The paper reports that this is particularly effective at low SNR and, in some low-SNR regimes, can even yield higher PSNR than the model without encryption/decryption because the super-resolution stage compensates for degradation in the latent representation (Zhang et al., 5 Sep 2025).

The super-resolution module is RDN-based and includes attention and residual mechanisms. Its described stages are:

1. mean subtraction preprocessing,

$s_{\text{trans}} = h * s_{\text{enc}} + n,$2

2. UNet-based attention map,

$s_{\text{trans}} = h * s_{\text{enc}} + n,$3

3. initial upsampling feature,

$s_{\text{trans}} = h * s_{\text{enc}} + n,$4

4. shallow feature extraction,

$s_{\text{trans}} = h * s_{\text{enc}} + n,$5

5. residual dense blocks,

$s_{\text{trans}} = h * s_{\text{enc}} + n,$6

with internal operations

$s_{\text{trans}} = h * s_{\text{enc}} + n,$7

$s_{\text{trans}} = h * s_{\text{enc}} + n,$8

$s_{\text{trans}} = h * s_{\text{enc}} + n,$9

6. global feature fusion,

$s_{\text{dec}} = Dec(s_{\text{trans}}),$0

7. attention modulation,

$s_{\text{dec}} = Dec(s_{\text{trans}}),$1

8. upsampling,

$s_{\text{dec}} = Dec(s_{\text{trans}}),$2

9. global residual learning,

$s_{\text{dec}} = Dec(s_{\text{trans}}),$3

10. final convolution and output,

$s_{\text{dec}} = Dec(s_{\text{trans}}),$4

$s_{\text{dec}} = Dec(s_{\text{trans}}),$5

Here, $s_{\text{dec}} = Dec(s_{\text{trans}}),$6 is ReLU, $s_{\text{dec}} = Dec(s_{\text{trans}}),$7 is the convolution weight in the $s_{\text{dec}} = Dec(s_{\text{trans}}),$8-th layer of the $s_{\text{dec}} = Dec(s_{\text{trans}}),$9-th residual dense block, and $\hat{y} = F_d(s_{\text{dec}}),$0 is the local feature fusion $\hat{y} = F_d(s_{\text{dec}}),$1 convolution. In aggregate, the module combines dense feature reuse, residual learning, and attention modulation to improve perceptual reconstruction.

5. Experimental configuration and evaluation protocol

The experimental study uses DIV2K for training and validation and 25 images from Urban100 for testing. All images are cropped to $\hat{y} = F_d(s_{\text{dec}}),$2 pixels. The implementation environment is Python 3.11.4 with PyTorch 2.5.1 on an NVIDIA RTX 6000 Ada Generation GPU. The network is trained at 10 dB SNR, and the loss is MSE (Zhang et al., 5 Sep 2025).

Component	Reported setting
Training/validation dataset	DIV2K
Test dataset	25 images from Urban100
Image size	$\hat{y} = F_d(s_{\text{dec}}),$3
Backbone	NTSCC
Training SNR	10 dB
Loss	MSE
Framework	Python 3.11.4, PyTorch 2.5.1
Hardware	NVIDIA RTX 6000 Ada Generation GPU

The hyperparameters are reported as 50 epochs, test every 1000 iterations, batch size 8, learning rate $\hat{y} = F_d(s_{\text{dec}}),$4, step decay factor $\hat{y} = F_d(s_{\text{dec}}),$5, Adam optimizer, Adam beta $\hat{y} = F_d(s_{\text{dec}}),$6, epsilon $\hat{y} = F_d(s_{\text{dec}}),$7, weight decay 0, and gradient clipping 0.

The physical channel is modeled as AWGN with complex Gaussian noise,

$\hat{y} = F_d(s_{\text{dec}}),$8

and the modulation schemes evaluated are 16QAM, QPSK, and BPSK. Robustness is assessed by varying SNR. The main quantitative metric is PSNR.

The paper also studies a scaling factor $\hat{y} = F_d(s_{\text{dec}}),$9 that maps semantic feature entropy to channel bandwidth cost. Larger $\hat{x} = G_s(\hat{y}),$0 corresponds to lower compression and more bandwidth for high-entropy patches, whereas smaller $\hat{x} = G_s(\hat{y}),$1 corresponds to higher compression and greater detail loss in high-entropy regions. This parameter is used to analyze the interaction between semantic compression and channel conditions.

The reported comparisons include NTSCC, NTSCC with the same encryption method, a baseline scheme without key-based decryption in the eavesdropping discussion, and variants with and without the super-resolution module.

6. Quantitative behavior, operating regimes, and stated significance

The reported PSNR–SNR curves and visual reconstructions show a clear modulation-dependent ordering under low SNR. The paper states that 16QAM has the lowest robustness at low SNR, QPSK has intermediate performance, and BPSK has the strongest noise resistance, with rapid PSNR improvement from 0–4 dB. The stated explanation is that low-order modulation is more noise tolerant, whereas high-order modulation has higher BER at the same SNR. Accordingly, BPSK gives the best low-SNR reconstruction quality, followed by QPSK and then 16QAM (Zhang et al., 5 Sep 2025).

Encryption and decryption are reported to amplify the impact of bit errors under low SNR, while having almost no effect on reconstruction quality at high SNR. The super-resolution module significantly improves PSNR under low SNR, especially for QPSK below 8 dB and BPSK below 5 dB. The benefit diminishes at high SNR and can become smaller than the non-SR case because inferred details may be less accurate than directly transmitted details when the channel is already clean.

At 4 dB SNR, the paper gives the following representative comparisons:

Scheme at 4 dB	Reported PSNR relation
16QAM SREC	similar to NTSCC; 2.5 dB higher than NTSCC with the same encryption method
QPSK SREC	3.0 dB higher than NTSCC; 4.6 dB higher than encrypted NTSCC
BPSK SREC	2.2 dB higher than NTSCC; 6.4 dB higher than encrypted NTSCC

The $\hat{x} = G_s(\hat{y}),$2-sensitivity study further differentiates operating regimes. Under 16QAM, PSNR remains lower across the examined range and grows slowly with $\hat{x} = G_s(\hat{y}),$3 because channel errors corrupt the additional transmitted bits. Under BPSK and QPSK, PSNR improves more sharply as $\hat{x} = G_s(\hat{y}),$4 increases but saturates after about $\hat{x} = G_s(\hat{y}),$5. For BPSK, the largest slope occurs near $\hat{x} = G_s(\hat{y}),$6, after which the gains diminish. The paper also notes that under good channel conditions, improving SNR matters less than avoiding information loss from compression.

Qualitatively, the visual results are described as showing more distortion for 16QAM under low SNR, better contour preservation for BPSK, sharper images and more visible detail with the super-resolution module, and the inability of an eavesdropper to recover meaningful image content without the key.

The paper’s broader significance is framed in secure semantic communication. Its main contributions are a secure semantic communication framework, modulo-256 encryption for semantic features, super-resolution enhancement via an RDN-based module, demonstrated low-SNR robustness under BPSK and QPSK relative to NTSCC and encrypted NTSCC baselines, and security-preserving semantic transmission. The stated implication is that semantic compression, encryption, and reconstruction enhancement can be jointly considered in a communication stack rather than optimized in isolation (Zhang et al., 5 Sep 2025).

7. Assumptions, limitations, and boundary conditions

SREC is explicitly developed under several assumptions. The transmitter and receiver share the encryption key in advance; the key is delivered through a secure channel; the eavesdropper may know all model parameters but not the key; the physical channel is AWGN; training is performed at 10 dB SNR; and the target task is image transmission (Zhang et al., 5 Sep 2025).

The paper also implies several limitations. Security depends on secure key distribution. The modulo-256 mechanism is lightweight, but the work does not provide a formal cryptographic proof beyond semantic-security intuition. The benefit of super-resolution declines at high SNR. High-order modulation suffers badly at low SNR. The study is limited to image transmission, Urban100 testing, and PSNR-based evaluation. More complex fading channels, adversarial channels, and tasks beyond image transmission are not analyzed. The effectiveness of the overall framework also depends on the quality of the underlying JSCC model.

Within those limits, SREC defines a concrete design point in secure SemCom: encrypted latent transmission combined with learned image enhancement. A plausible implication is that future secure semantic communication systems may be evaluated not only by confidentiality and channel robustness separately, but by how successfully they co-design semantic representation, protection of latent features, and task-aware reconstruction.