HANA-JSCC: Robust MIMO Semantic Transmission

• HANA-JSCC is a semantic image transmission system that robustly adapts to imperfect CSI in MIMO channels using a Transformer-based channel matrix adaptor.
• It integrates a Swin-Transformer semantic encoder/decoder with a split channel codec and CSI-conditioned processing to refine latent representations.
• The approach employs a two-stage training with knowledge distillation, yielding 0.40–0.54 dB PSNR improvements over state-of-the-art benchmarks on datasets like CIFAR-10 and Kodak24.

Searching arXiv for the HANA-JSCC paper and closely related semantic communication context. HANA-JSCC is a semantic image transmission system with channel matrix and channel noise adaptation for multiple-input-multiple-output semantic communication under imperfect channel state information (CSI). It was introduced to address the impracticality of assuming perfect channel matrix estimation in channel-adaptive joint source-channel coding, particularly under hardware and pilot overhead constraints. The method combines a Swin-Transformer-based semantic encoder and decoder, a split channel codec, a Transformer-based channel matrix adaptor, and a two-stage training procedure with knowledge distillation. In the reported evaluations, it improves average image reconstruction quality over state-of-the-art benchmarks across multiple noise and estimation-error settings, with gains of approximately $0.40\sim0.54$ dB in average PSNR depending on dataset and condition (Gong et al., 4 Sep 2025).

1. Problem setting and formal model

HANA-JSCC is situated in semantic communication for image transmission over MIMO channels. The motivating problem is that existing MIMO semantic communication systems assume perfect channel matrix estimation for channel-adaptive joint source-channel coding, whereas the target deployment regime involves imperfect CSI at both transmitter and receiver. The system therefore treats channel estimation error as a first-class design variable rather than as a negligible perturbation (Gong et al., 4 Sep 2025).

The channel model is a flat Rayleigh fading MIMO channel with imperfect CSI. The true channel matrix is

$H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$

and the estimated channel available at both ends is

$$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$

After precoding, the transmitted symbols satisfy

$$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$

with

$N\sim\mathcal{CN}(0,\sigma_n^2 I).$

Receiver-side post-processing with the estimated left singular vectors yields

$$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$

This formulation makes the mismatch between $H_{est}$ and $H_p$ central to the decoding problem. A plausible implication is that any end-to-end architecture designed only for perfect CSI will be structurally vulnerable to misalignment introduced by estimation noise, even if its source and channel representations are otherwise strong.

2. End-to-end architecture

The architecture is organized as a transmitter, a MIMO channel, and a receiver, with explicit CSI-conditioned processing inserted into the latent channel-coding path (Gong et al., 4 Sep 2025).

At the transmitter, an input image $x$ is first mapped by a semantic encoder $E_S(\cdot\mid\phi_S)$, described as a Swin-Transformer-based network, into a lower-dimensional semantic feature map

$H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$0

The encoder also takes the scalar channel SNR $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$1 as side input through a small MLP denoted Channel-ModNet. The channel codec is split into two MLP sub-codecs:

• $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$2
• $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$3

A non-trainable power-normalization layer follows $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$4. Between these two sub-codecs sits the channel matrix adaptor $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$5, which refines the latent representation using $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$6. The resulting complex symbols are then precoded by $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$7, obtained from the singular value decomposition of the estimated channel matrix.

The receiver mirrors this logic. Post-processing with $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$8 produces $H_p \in \mathbb{C}^{n_{rx}\times n_{tx}},$9, which is passed through a three-part channel decoder:

• $$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$0
• $$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$1
• $$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$2

A semantic decoder $$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$3, also with SNR attention, reconstructs the image $$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$4.

The corresponding end-to-end equations are

$$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$5

$$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$6

$$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$7

and, at the receiver,

$$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$8

$$H_{est} = H_p + H_e,\quad H_e\sim\mathcal{CN}(0,\sigma_e^2).$$9

$$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$0

3. Channel matrix adaptor

The distinguishing architectural element is the channel matrix adaptor, inserted symmetrically in the transmitter and receiver latent pipelines. Its role is to collaborate with the channel codec to adapt to misaligned channel state information, thereby mitigating the impact of estimation errors (Gong et al., 4 Sep 2025).

The adaptor takes two inputs: a latent token sequence

$$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$1

and the estimated channel matrix

$$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$2

The estimated CSI is projected by a small fully connected layer into a real CSI token

$$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$3

This token is concatenated with the latent sequence,

$$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$4

and processed by $$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$5 Transformer blocks, with $$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$6 in the reported configuration. Each block contains multi-head self-attention with $$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$7 heads of size $$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$8, followed by an MLP with hidden size $$y = H_p\,V_{est}\,z_C + N \in\mathbb{C}^{n_{rx}\times d},$$9, GeLU activation, residual connections, and layer normalization. The first token is discarded at the output, leaving $N\sim\mathcal{CN}(0,\sigma_n^2 I).$0 refined tokens $N\sim\mathcal{CN}(0,\sigma_n^2 I).$1, identified as $N\sim\mathcal{CN}(0,\sigma_n^2 I).$2.

The adaptor transformations are written as

$N\sim\mathcal{CN}(0,\sigma_n^2 I).$3

and

$N\sim\mathcal{CN}(0,\sigma_n^2 I).$4

The paper attributes the need for this design to an ill-posed one-to-many relation between estimated and true channel matrices. By explicitly embedding the estimated CSI into a learnable token and using self-attention over CSI and semantic tokens, the adaptor is described as learning a many-to-one mapping that aligns the latent representation with the actual channel realization. It does not attempt to recover a unique true channel from $N\sim\mathcal{CN}(0,\sigma_n^2 I).$5; rather, it adjusts feature tokens conditioned on the observed estimate. This directly addresses a common misconception that robustness to imperfect CSI must come from explicit channel reconstruction alone.

4. Two-stage training and knowledge distillation

HANA-JSCC uses a two-stage training strategy designed to overcome convergence difficulties caused by the ill-posed problem. The first stage performs imperfect-CSI fine-tuning, and the second stage introduces knowledge distillation from a perfect-CSI teacher to an imperfect-CSI student (Gong et al., 4 Sep 2025).

In Stage I, the entire network is initialized from a SwinJSCC model pretrained under imperfect CSI. The semantic encoder and decoder parameters $N\sim\mathcal{CN}(0,\sigma_n^2 I).$6 are frozen, while the channel codec and adaptor parameters

$N\sim\mathcal{CN}(0,\sigma_n^2 I).$7

are trained. The loss is the image-domain $N\sim\mathcal{CN}(0,\sigma_n^2 I).$8 objective

$N\sim\mathcal{CN}(0,\sigma_n^2 I).$9

with $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$0 sampled uniformly from $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$1 in each batch.

In Stage II, the teacher is a copy of the same architecture trained under perfect-CSI conditions in Stage I, with all weights frozen. The student is the Stage-I imperfect-CSI model, now fully unfrozen. Distillation is applied through KL-divergence at two internal levels: the post-adaptor channel token $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$2 and the recovered semantic feature $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$3. The Stage-II loss is

$$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$4

The reported training setup typically sets $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$5, doubles the batch size, and uses Adam with learning rate $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$6 and cosine decay.

This training strategy is technically notable because distillation is imposed on intermediate latent variables as well as the final reconstruction objective. This suggests that the method treats imperfect-CSI robustness as a representation-alignment problem, not solely as an output-level denoising problem.

5. Experimental configuration and reported results

The reported evaluation uses two datasets and a fixed large-array MIMO configuration. CIFAR-10 is used with $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$7k training images and $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$8k test images of size $$\tilde z_C = U_{est}^H\,y = U_{est}^H H_p V_{est} z_C + U_{est}^H N.$$9, at compression ratio $H_{est}$0. Kodak24 is evaluated on $H_{est}$1 high-resolution images of size $H_{est}$2 or $H_{est}$3, at compression ratio $H_{est}$4, with training on $H_{est}$5k ImageNet crops. The MIMO setting is $H_{est}$6. Channel noise SNR is sampled from $H_{est}$7 dB, and the estimation error variance is uniformly distributed in $H_{est}$8 (Gong et al., 4 Sep 2025).

The benchmarks are DeepJSCC, DeepJSCC-V, DeepJSCC-MIMO, SwinJSCC, and SwinJSCC-w/o-SA. All are pretrained with perfect CSI and then fine-tuned on $H_{est}$9.

Dataset / metric	Method	Reported result
CIFAR-10 average over all SNR and $H_p$0	DeepJSCC-MIMO	28.14 dB
CIFAR-10 average over all SNR and $H_p$1	SwinJSCC	28.09 dB
CIFAR-10 average over all SNR and $H_p$2	HANA-JSCC	28.54 dB
Kodak24 average	SwinJSCC	30.20 dB
Kodak24 average	HANA-JSCC w/o KD	30.62 dB
Kodak24 average	HANA-JSCC w/ KD	30.74 dB

On CIFAR-10, the gain is reported as approximately $H_p$3 to $H_p$4 dB over the best listed baselines. On Kodak24, the adaptor contributes $H_p$5 dB relative to SwinJSCC, and knowledge distillation contributes a further $H_p$6 dB. The broader summary given for the method is that HANA-JSCC achieves $H_p$7 dB higher average performance across various noise and estimation error levels in various datasets.

The reported discussion further states that, across the full SNR range and $H_p$8 span, HANA-JSCC consistently outperforms all baselines by approximately $H_p$9–$x$0 dB PSNR on CIFAR-10 and approximately $x$1 dB on Kodak24. It also states that the Transformer-based adaptor absorbs channel estimation errors and restores the channel decomposition into approximately parallel sub-channels even under misaligned CSI.

6. Interpretation, scope, and relation to neighboring approaches

Within the reported design space, HANA-JSCC combines three elements: a two-step channel codec with an in-the-loop CSI adaptor, a lightweight Transformer design for feature refinement, and two-stage training with KL-based knowledge distillation (Gong et al., 4 Sep 2025). The system therefore differs from benchmarked approaches not by replacing semantic image coding altogether, but by inserting an explicit CSI-conditioned latent adaptation mechanism and pairing it with a teacher-student training scheme.

Its scope is specifically semantic image transmission over large MIMO arrays with imperfect CSI. The paper does not claim generality beyond the reported flat Rayleigh fading MIMO setting, the specified dataset configurations, and the $x$2 experiments. A plausible implication is that the method should be understood as a robustification of MIMO semantic communication under channel-estimation mismatch, rather than as a universal solution for arbitrary channel models or modalities.

An important conceptual point is that the method does not remove the use of SVD-based precoding and post-processing. Instead, it augments that pipeline so that latent variables can adapt when the estimated decomposition is misaligned with the true channel. Likewise, the knowledge distillation stage does not substitute for imperfect-CSI training; it is added after Stage-I fine-tuning and is reported to be especially beneficial in high-SNR and low-error regimes.

In summary, HANA-JSCC is a MIMO semantic image transmission framework that operationalizes robustness to imperfect CSI through CSI-tokenized latent refinement and dual-level knowledge distillation. In the reported experiments, this combination yields consistently higher PSNR than the listed state-of-the-art baselines while preserving an end-to-end differentiable semantic communication architecture.