Importance-Aware Robust Semantic Transmission

• IRST is a framework that prioritizes task-critical semantic features by dynamically estimating importance from deep models.
• It employs weighted source-channel coding and cross-layer adaptive strategies to allocate power, rate, and modulation based on semantic relevance.
• Empirical studies demonstrate significant gains in accuracy, rate efficiency, and resilience against packet loss and adversarial impairments compared to traditional methods.

Importance-Aware Robust Semantic Transmission (IRST) denotes a class of semantic communication methods that allocate transmission, coding, and reconstruction effort according to the task relevance of transmitted content, rather than treating all bits, pixels, features, or packets as equally valuable. In the cited literature, IRST appears as weighted deep joint source-channel coding for wireless image transmission (Sun et al., 2023), as cross-layer semantic importance-aware communications using pre-trained LLMs (Guo et al., 2023), as robust semantic image transmission under adversarial semantic impairment (Peng et al., 2024), as understanding-level semantic communications with LLM-based semantic correction (Guo et al., 2024), as adaptive feature scheduling over time, space, and space-time (Zhou et al., 2023, Zhou et al., 2024), as packet-loss-resilient representation design (Yang et al., 21 Nov 2025), as task-oriented rate control (Sun et al., 29 Apr 2025), as semantic-aware HARQ (Hu et al., 2024), and as satellite-ground, transport-layer, split-interface, and semantic-aware physical-layer designs (Cao et al., 15 Aug 2025, Wang et al., 28 Apr 2026, Kong et al., 17 Jul 2025, Shaju et al., 14 May 2026). Across these variants, the common objective is to preserve semantic fidelity and downstream task performance under constrained bandwidth, time-varying channels, packet loss, or semantic impairment.

1. Conceptual basis and problem formulation

IRST departs from source-channel separation and uniform-fidelity transmission by assuming that semantic representations are heterogeneous: some components contribute disproportionately to classification, detection, reconstruction, or control. In the image-transmission formulation of "Deep Joint Source-Channel Coding for Wireless Image Transmission with Semantic Importance" (Sun et al., 2023), the central design choice is to preserve semantic information during wireless image transmission so as to boost the performance of intelligent tasks for images at the receiver. The method first computes semantic importance weights from a fixed downstream task network, then uses those weights inside a semantic loss, and finally trains a deep JSCC encoder-decoder end-to-end with that loss.

A representative feature-level objective is the weighted feature reconstruction loss

$$\mathcal{L}_{\text{SD-JSCC}}=\frac{1}{B}\sum_{b=1}^{B}\sum_{k=1}^{K} w_k' \cdot \|f_k^{b'}-f_k^b\|^2,$$

where the weighting vector $W'=\{w_k'\}$ encodes semantic importance (Sun et al., 2023). This formulation makes the target of communication the preservation of task-relevant latent features rather than pixelwise fidelity alone.

A parallel line of work reframes semantic transmission at the understanding level. In "Semantic Importance-Aware Communications with Semantic Correction Using LLMs" (Guo et al., 2024), an image caption neural network converts visual data into natural language descriptions, LLMs quantify frame importance, and semantic loss is written as

$$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$

The same framework uses LLM-based semantic correction at the receiver and, if desired, text-to-image generation for visual data regeneration. This establishes that IRST is not restricted to feature tensors: it also includes natural-language semantic units.

The literature therefore presents IRST as a general task-oriented principle: semantic fidelity is optimized through unequal protection, unequal resource allocation, or unequal reconstruction effort, with robustness defined relative to the impairments most damaging to the intended task.

2. Estimating semantic importance

Importance estimation is the defining step in IRST, and the literature spans multiple granularities, from feature maps to words, patches, segments, bits, and latent symbols.

Granularity	Importance definition	Representative systems
Feature maps	Gradient of task output or loss with respect to features	SD-JSCC, FAST, IRCSC, SemHARQ
Frames or words	Important-word counts or omission-induced semantic loss	SIAC, ULSC
Patches or regions	Object regions, SIFT-ranked patches, or dynamic patch masks	ASCViT-JSCC, DeepSC-RI
Segments, tokens, bits	Segment-level, receiver-specific, token-level, and bit-level importance	Multi-modal task-oriented SemCom, split-interface JSCC
Latent concepts or symbols	Semantic criticality indicator over discrete latent concepts	Semantic-aware constellation design

In the gradient-based semantic importance weight module of (Sun et al., 2023), the weight of the $k$-th feature map for perception result $y^c$ is

$$w_k^c=\frac{1}{M\times N}\sum_{m=1}^{M}\sum_{n=1}^{N}\frac{\partial y^c}{\partial f_k}, \qquad w_k=\frac{1}{C}\sum_{c=1}^{C} w_k^c,$$

followed by temperature-scaled softmax normalization,

$W' = r \cdot \mathrm{Softmax}(\tau \times W).$

Related gradient-based ranking appears in FAST, where feature importance is derived from the average gradient of the system loss with respect to each feature (Zhou et al., 2023), in IRCSC, where channel importance is computed from the gradient of the predicted probability for the true label with respect to each feature channel (Sun et al., 29 Apr 2025), and in SemHARQ, where feature importance ranking is based on gradients of multi-task outputs (Hu et al., 2024).

Text-centric schemes quantify semantic importance explicitly through LLMs. In SIAC, ChatGPT-SIAC counts important words, while BERT-SIAC defines importance through the semantic loss caused by omitting a word or frame, using cosine similarity in the BERT embedding space (Guo et al., 2023). ULSC uses LLM completion to evaluate how difficult it is to recover a removed frame:

$w_l = 1-\phi(\mathbf{m}_0,\mathbf{m}_l),$

where $\phi$ is the BERT-based semantic similarity between the original and completed sentence (Guo et al., 2024).

Image-region and patch-wise importance is also common. ASCViT-JSCC identifies object regions with YOLOv5 and ranks non-object patches by SIFT feature points before adaptive masking (Ding et al., 2024). DeepSC-RI uses a semantic importance evaluation module to assign scores to image patches and suppress the least important patches in self-attention through a dynamic mask $\mathbf{M}_d$ (Peng et al., 2024).

In multi-modal task-oriented semantic communication, importance is formalized at several levels: Segment-Level Importance (SLI), Receiver-Specific Importance (RSI), Token-Level Importance (TLI), and Bit-Level Importance (BLI) (Ma et al., 22 Feb 2025). The split-interface design of (Kong et al., 17 Jul 2025) further treats bit-flip probabilities $W'=\{w_k'\}$0 in per-bit binary symmetric channels as trainable indicators of bit-level importance, and the Importance-Aware Net consumes that interface-derived importance to guide channel mapping and demapping.

At the physical layer, "Not All Symbols Are Equal" (Shaju et al., 14 May 2026) introduces a semantic criticality indicator (SCI) that scores each discrete latent concept by task relevance. This extends importance awareness from source representation to constellation geometry.

3. Resource allocation and adaptive transmission policies

Once importance is estimated, IRST uses it to control power, rate, coding depth, scheduling, or precoding. The result is a cross-layer family of unequal-protection mechanisms.

SIAC formulates semantic importance-aware power allocation over Rayleigh fading channels. With frame importance weights $W'=\{w_k'\}$1 and per-frame outage probability

$W'=\{w_k'\}$2

the design minimizes the expected semantic importance-weighted outage

$W'=\{w_k'\}$3

This framework is explicitly intended for direct embedding into current communication systems through a cross-layer manager rather than through end-to-end retraining of a Deep-JSCC pipeline (Guo et al., 2023).

ULSC performs semantic-aware adaptive modulation and coding by minimizing semantic loss subject to a delay constraint,

$W'=\{w_k'\}$4

with

$W'=\{w_k'\}$5

A greedy search assigns modulation/coding pairs to frames according to their semantic importance (Guo et al., 2024).

IRCSC treats adaptive rate control as the selection of the minimum number of feature channels $W'=\{w_k'\}$6 satisfying a semantic transmission integrity requirement,

$W'=\{w_k'\}$7

where $W'=\{w_k'\}$8 is the Semantic Transmission Integrity Index (STII). Because $W'=\{w_k'\}$9 increases monotonically with $$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$0, the paper uses binary search to find the minimum admissible rate (Sun et al., 29 Apr 2025).

Feature scheduling in time-varying channels appears in FAST. The feature priority metric combines importance and robustness,

$$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$1

and the transmitter predicts future CSI, sorts both features and channel states, and transmits the prior features under better CSI (Zhou et al., 2023). The later space-time FAST framework generalizes the same idea to SISO and MIMO settings, using predicted future CSI for temporal allocation and SINR or singular values for space-time allocation without intricate fine-tuning (Zhou et al., 2024).

Satellite-ground IRST uses channel-aware semantic selection and adaptive coding under rapid SNR variation and bandwidth scarcity. The framework first improves segmentation granularity with a segmentation model enhancement algorithm, then invokes a task-driven semantic selection method based on real-time channel state information, and finally applies a stack-based, SNR-aware channel codec with progressively activated coding depth (Cao et al., 15 Aug 2025).

Foundation-model-based adaptive semantic image transmission likewise uses task-adaptive precoding. After decomposing an image into a semantic segmentation map and a compressed representation, the system applies a trainable weighting matrix and SVD-based channel decomposition so that the best subchannels carry the most important semantic features (Liu et al., 28 Sep 2025).

4. Robustness mechanisms beyond additive noise

A distinctive feature of IRST is that robustness is not limited to AWGN or fading. The literature treats semantic impairment, adversarial perturbation, packet loss, retransmission control, and transport-layer header failure as first-class problems.

DeepSC-RI focuses on semantic impairments in images, particularly those arising from adversarial perturbations. It introduces the Image Semantic Impairment Intensity (ISII),

$$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$2

constructs a semantic impairment dataset on CIFAR10 using PGD attacks, and uses a dual-branch multi-scale semantic extractor. The fine-grained branch emphasizes crucial patches via a semantic importance evaluation module and dynamic masking, while the coarse-grained branch captures robust semantics hierarchically; the two are fused with cross-attention (Peng et al., 2024).

Packetized semantic communication reveals a different failure mode: the loss of any packet equates to the complete loss of all semantic information it carries. "Feature Partitioning and Semantic Equalization for Intrinsic Robustness in Semantic Communication under Packet Loss" (Yang et al., 21 Nov 2025) shows that both Transformer- and CNN-based systems are most robust when features are partitioned along the channel dimension. The paper further argues that balanced semantic representation is fundamental to intrinsic robustness. For CNNs, the proposed Semantic Equalization Mechanism (SEM) combines a Dynamic Scale module and a Broadcast module to prevent a few channels from dominating.

SemHARQ addresses robustness through selective retransmission. Its Feature Importance Ranking (FIR) selects the most task-relevant features for initial transmission, while a Feature Distortion Evaluation (FDE) network estimates post-channel distortion at the receiver. Corrupted features are retransmitted, and remaining channel resources are used for incremental transmissions. The framework targets multi-task scenarios in Internet of Vehicles and couples retransmission policy to semantic uncertainty (Hu et al., 2024).

Transport robustness is taken up by SPAT. Instead of explicit source and destination port headers, SPAT embeds port information into semantic representations:

$$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$3

where the bias is derived from learnable source and destination port embeddings. Uplink uses semantic port identification, downlink uses destination-aware conditional gating, and an adaptive-rate controller varies the number of transmitted semantic channels according to channel conditions and feature importance (Wang et al., 28 Apr 2026). This directly addresses vulnerability to header corruption and resulting packet loss.

A plausible implication of these works is that IRST has become a multi-impairment framework: the object of protection may be semantic content, semantic routing metadata, or balanced latent structure, depending on where the dominant failure occurs.

5. Modalities and architectural realizations

The IRST literature is modality-diverse. Wireless image transmission remains the most developed setting, but point clouds, multi-modal and multicast traffic, satellite-ground imagery, and semantic-aware modulation are all represented.

For images, weighted feature-space Deep-JSCC constitutes an early feature-level realization (Sun et al., 2023). Later image systems emphasize architectural specialization. ASCViT-JSCC integrates ViTs with OFDM, quantization, YOLOv5-based object detection, SIFT-based feature-point ranking, and a masked autoencoder for recovery, with over-the-air validation on an SDR-based prototype (Ding et al., 2024). Foundation-model-based image transmission decomposes each image into a semantic segmentation map and a compressed representation, estimates channel knowledge with a conditional diffusion model, and reconstructs with ControlNet-guided Stable Diffusion (Liu et al., 28 Sep 2025).

Point-cloud IRST is represented by the Point-BERT-based semantic-aware transmission system of (Han et al., 2023). The semantic encoder uses Farthest Point Sampling, kNN grouping, lightweight local feature extraction, and a pretrained Point-BERT transformer to produce latent semantic representations for robust classification. A two-stage training strategy first learns the semantic encoder/decoder on clean data and then trains the full pipeline with random SNRs from 0 dB to 20 dB, while freezing the pretrained transformer and positional embeddings in stage two.

Multi-modal task-oriented semantic communication extends importance awareness to semantic segments, tokens, and bits. GenAI partitions visual data objects into semantic segments, encodes them into tokens, and supports importance-aware source and channel coding across multicast scenarios in which segment importance varies among receivers. Rate-splitting coded progressive transmission is proposed to ensure flexibility and robustness (Ma et al., 22 Feb 2025).

Satellite-ground IRST treats remote sensing or scene imagery under bandwidth scarcity and rapidly varying SNR. The dedicated framework in (Cao et al., 15 Aug 2025) combines segmentation enhancement, task-driven semantic selection, and an SNR-aware stacked codec tailored to low Earth orbit channels.

At the interface and physical layers, IRST also appears in decoupled designs. The learning-based interface of (Kong et al., 17 Jul 2025) preserves the split between application-layer source coding and lower-layer channel coding while exposing bit-level importance through trainable interface parameters. At the modulation stage, (Shaju et al., 14 May 2026) uses a VQ-VAE to extract discrete latent concepts, SCI to rank them, a deep reinforcement learning agent to select the transmission subset under instantaneous channel conditions, and a learned semantic-aware $$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$4-QAM constellation to protect critical concepts physically.

This architectural diversity indicates that IRST is not a single codec family. It is a design principle that can be instantiated in end-to-end neural JSCC, cross-layer digital systems, retransmission protocols, semantic transport, or semantic-aware modulation.

6. Metrics, empirical findings, and research directions

Evaluation in IRST is explicitly multi-metric. The literature uses downstream task metrics such as ACC, F1-score, mAP, classification accuracy, IoU, and rank-1 accuracy; perceptual or distortion metrics such as PSNR, SSIM, LPIPS, and FID; and new semantic metrics such as semantic loss, STII, ISII, Semantic Symbol Vulnerability (SSV), and Semantic Protection Probability (SPP) (Sun et al., 2023, Peng et al., 2024, Sun et al., 29 Apr 2025, Shaju et al., 14 May 2026).

Quantitative gains are often substantial. In weighted Deep-JSCC for wireless images, the proposed method achieves up to 57.7% and 9.1% improvement in intelligent task performance compared with source-channel separation coding and deep source-channel joint coding without semantics, respectively, at the same compression rate and SNR; on ImageNet at $$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$5 bpp it reaches 82.16% ACC versus 71.22% for deep JSCC and 79.57% for BPG, and on Pascal VOC at $$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$6 bpp and $$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$7 dB it achieves 51.11% mAP versus 45.79% for deep JSCC (Sun et al., 2023). In robust point-cloud classification, the Point-BERT-based system attains classification accuracy of over 89% when SNR is higher than 10 dB and still maintains accuracy above 66.6% at 4 dB, outperforming the baseline by 0.8% to 48% across SNR values (Han et al., 2023). Under packet loss, CNNs equipped with SEM retain about 85% of lossless PSNR at 40% packet loss (Yang et al., 21 Nov 2025). In ULSC, semantic similarity remains in 0.82–0.84 at a 10% frame erasure rate, versus 0.65 or lower for DeepJSCC, while attacker reconstruction quality drops to PSNR $$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$8–$$SL=\sum_{l=1}^{L} w_l P_{l,\text{frame}}^{\text{Loss}}.$$9 dB versus $k$0–$k$1 dB for DeepJSCC (Guo et al., 2024). IRCSC reduces average rate to about 14.48 kbps in AWGN and 20.61 kbps in Rayleigh at 12 dB, compared with 125.44 kbps for TD-JSCC and 47.04 kbps for WO-FS (Sun et al., 29 Apr 2025). SemHARQ reports more than 20% improvement in rank-1 accuracy for vehicle re-identification and 10% in vehicle color classification accuracy in the low-SNR regime (Hu et al., 2024). SPAT achieves PSNR of approximately 26.8 dB in real-world experiments while maintaining low-latency transmission (Wang et al., 28 Apr 2026). Semantic-aware constellation design reports near 100% SPP across modulation orders from 4-QAM to 1024-QAM versus 50% for standard constellations at high spectral efficiency, together with a 21:1 compression ratio and semantic quality above 0.9 (Shaju et al., 14 May 2026).

Several research tensions are explicit in the literature. One is the contrast between feature-level semantic communications and understanding-level semantic communications; the latter claim superior ability to convey semantic understanding and improved privacy because neither original data nor features are directly transmitted (Guo et al., 2024). A second is deployment architecture: end-to-end Deep-JSCC yields tightly coupled source-channel designs, whereas SIAC and split-interface schemes emphasize compatibility with existing systems through cross-layer managers or digital interfaces (Guo et al., 2023, Kong et al., 17 Jul 2025). A third concerns robustness itself: packet-loss studies argue that balanced channel utilization is a fundamental condition for intrinsic robustness, whereas other lines of work focus on selecting and heavily protecting only a sparse set of critical features (Yang et al., 21 Nov 2025). A fourth concerns evaluation: STII, ISII, CLIP similarity, LPIPS, IoU, SSV, and SPP were introduced precisely because BER or PSNR alone do not capture semantic integrity (Peng et al., 2024, Sun et al., 29 Apr 2025, Guo et al., 2024, Shaju et al., 14 May 2026).

Taken together, these works portray IRST as a cross-layer and cross-modal research program centered on one proposition: transmission systems should not only compress meaning, but also identify which meaning matters most, anticipate how it may be damaged by the channel or protocol stack, and allocate protection accordingly.