{"id":"b7dda0cc-078b-401a-ba07-ff4fb1219ede","arxiv_id":"2412.07173","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A method for joint transmission of images and digital signals by mapping binary data to mask locations in a Masked Autoencoder, with a sparse index encoding to reduce overhead.","lead":"This paper proposes hiding digital data in the masked-out patches of an image and using a pre-trained Masked Autoencoder to transmit image and data together. It reports good image quality and bit-error rates in simulation, but key mapping and channel details are left unspecified.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The digital-data path is not a working system as written: Algorithm 1/2 can invert the recovered bit stream (no flag for whether the sparse matrix holds masked or unmasked indices), and no channel scheme for the index list is specified, so the reported BER is not tied to a concrete transmission.","rationale":"The reader rejected the paper because the bit-to-mask mapping is never described and the reported BER is not tied to a concrete transmission scheme. My stress-test confirms this and sharpens it to an internal inconsistency: as written, the sparse-matrix encoder chooses the minority class without telling the receiver which class it chose, so the decoder can reconstruct the complement of the transmitted mask. This is not a missing optimization or a mere ambiguity; it is a definitive correctness break for a valid input (e.g., a bit string with more ones than zeros). The additional undefined `ids restore` and the spurious '-1' in the length computation reinforce that the algorithmic description is not self-contained. The overhead-reduction argument in Section III-C also depends on Lm being small, but the paper never specifies how the sparse index list is represented or transmitted; if it is sent as a list of integer indices, the cost is Lb·log2(N), which must be weighed against the savings from masking Lb patches. The simulation results are therefore not reproducible without substantial unstated design choices. These issues are load-bearing because the entire digital-signal contribution is carried by the mask pattern; if the mask recovery is not provably invertible and its transmission cost is not accounted for, the paper's central claim is unsupported. I find no independent evidence, such as code or proofs, that would mitigate these gaps. The reader's reject verdict remains appropriate, so no adjustment is needed.","tokens_in":14902,"tokens_out":10078,"duration_ms":110627,"concrete_test":"Implement the encoder/decoder pair exactly as in Algorithms 1 and 2 for a 224×224 image (N=196) and a 150-bit all-ones stream (so the minority class is zeros). Verify whether the receiver's recovered mask equals the transmitted mask. If it equals the complement, the decoder is not invertible. As a second check, count the bits required to transmit the sparse index list under any standard code (e.g., Lb·ceil(log2 N) for the list) and compare with the claimed savings (N−Lb)·Le from Section III-C; if the index cost is not negligible, the overhead-reduction claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim 'joint transmission without significant communication overheads' hinges on a concrete, invertible bit-to-mask mapping and on a specified way to send the mask indices. Section III-A says only that masked patch locations are 'determined by the indices of \"1\"s in the binary sequence,' without stating how a binary stream of length Lb is converted to an N-bit mask (padding? combinatorial code? length prefix?). Section III-B's Algorithms 1 and 2 are not invertible as written: Algorithm 1 chooses the sparse matrix as the set of indices of the *minority* class among masked/unmasked patches (lines 7-11), but Algorithm 2 unconditionally sets mask[:, sparse_matrix] = 1 (line 5). If the transmitted bit string has more ones than zeros, the sparse matrix contains the zero indices, and the receiver reconstructs the complement of the true mask, yielding a bitwise-inverted digital signal. The pseudocode also references the undefined `ids restore` variable and a spurious '-1' in `L <- Lm + Length(gen(Sv_im)) - 1` (Alg. 2, line 3), so the length recovery is incoherent. Independently, no modulation/channel-coding scheme for the sparse index list is described; because a single index error can corrupt the entire mask ordering via the argsort operations, the BER curves in Fig. 4 are not reproducible from the text. The digital payload rides entirely on the mask pattern, so these gaps are not cosmetic: without an explicit bijective code and error protection, the central claim of reliable digital-signal transmission is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a semantic-communication scheme in which a binary digital signal is mapped into the positions of masked patches of a carrier image, and a pre-trained masked autoencoder (MAE) is used to reconstruct the image while the mask pattern carries the digital payload. To reduce the overhead of sending the mask pattern, it adds a sparse encoding module that transmits the minority index set together with the latent representation of the unmasked patches. Simulations report PSNR/MS-SSIM for the image and BER for the binary data over AWGN and Rayleigh channels, and tables compare transmission bits and computation time against DeepJSCC-V, SwinJSCC, and a traditional MAE baseline.","tokens_in":15254,"tokens_out":7610,"duration_ms":80460,"significance":"The idea of embedding digital bits in the mask pattern of an MAE is original and potentially attractive, since MAE is designed to tolerate varying mask patterns and using a pre-trained model avoids rebuilding a semantic codec. The overhead comparison in Table I and the computation-time study in Table II are also useful as starting points. However, the central claim of reliable joint transmission without significant overhead is not established: the bit-to-mask mapping is not defined, the encoding and decoding algorithms are not inverses of each other in the high-mask-ratio regime that the paper emphasizes, and no physical-layer transmission scheme for the sparse index list is given. These are load-bearing gaps, not merely presentation issues.","major_comments":[{"comment":"The central assertion that the scheme achieves joint transmission 'without incurring significant communication overheads' (Abstract) and 'without additional communication costs' (Section III.A) is contradicted by Section III.B, which states that the original MAE 'still carries a significant overhead in transmitting M' and introduces the sparse encoding module 'to reduce the transmission costs.' The paper cannot simultaneously claim no overhead and then acknowledge that the mask tokens are a significant overhead; this internal inconsistency is load-bearing because overhead reduction is one of the stated contributions.","section":"Abstract; Section III.A; Section III.B"},{"comment":"The bit-to-mask mapping is never specified. The text says only that 'masked patch locations are determined by the indices of \"1\"s in the binary sequence,' and then immediately says 'the \"1\"s represent unmasked patches,' which is ambiguous. It does not explain how an arbitrary binary stream of length Lb is converted into the N-bit mask (e.g., fixed-length blocks, padding, or a combinatorial code), nor how the receiver knows Lb or the total number of patches. Since the entire digital payload is carried by this mapping, the BER results in Fig. 4 cannot be evaluated or reproduced from the description.","section":"Section III.A"},{"comment":"Algorithms 1 and 2 are not mutually inverse. Algorithm 1 (lines 7–11) sets sparse_matrix to the complement of the larger of the two sets ids_0 and ids_1, i.e., it transmits whichever class—masked or unmasked—is the minority. Algorithm 2 (line 5) unconditionally assigns 1 to the columns sparse_matrix, treating it as the masked set. When the masked class is the majority (the high-mask-ratio regime that the paper emphasizes), Algorithm 1 transmits the unmasked indices, and Algorithm 2 then marks the unmasked patches as masked, yielding the complement of the transmitted mask and hence a bitwise-inverted digital signal. No flag or side information distinguishes the two cases. In addition, Algorithm 1 uses the undefined variable `ids_restore` and a spurious `-1` in `len_keep`, and Algorithm 2 repeats the same `-1` in computing L, so the length recovery is incoherent.","section":"Section III.B, Algorithms 1 and 2"},{"comment":"No transmission scheme is specified for the sparse index list or the latent representation. The system model in Section II describes a complex baseband channel (Eq. (2)), but the simulations give no modulation, constellation, channel coding, or error protection for the transmitted indices and latent vectors, nor how an SNR value translates into discrete index errors. This matters because the receiver's argsort operations are globally sensitive: a single erroneous index can reorder the entire mask and destroy all payload bits. As written, the BER curves in Fig. 4 are not reproducible and are not connected to the claimed physical-layer model.","section":"Section IV.A and Fig. 4"},{"comment":"The overhead expression is dimensionally inconsistent and unjustified. If Le and Lm denote total numbers of bits for gen(Sv_im) and M, then multiplying Le by the number of unmasked patches is incorrect; if they are per-patch or per-token quantities, that is not stated. The subsequent simplification assumes Lb = Mr N and treats Lm as fixed, then concludes that adding more binary data reduces the total transmitted bits, but no analysis shows how Lm scales with the mask pattern or what happens when Lb exceeds N. Since overhead reduction is a headline claim, this analysis must be made precise.","section":"Section III.C"}],"minor_comments":[{"comment":"The symbol M is used to mean both the mask tokens and the masked-patch indices; these are different objects in an MAE decoder, and the paper should disambiguate them.","section":"Throughout"},{"comment":"The output M in Algorithm 2 is an index array obtained by argsort, but the text calls it 'recovered mask tokens'; the paper does not explain how the MAE decoder obtains the learned mask-token embeddings at the recovered indices.","section":"Algorithm 2"},{"comment":"The baseline name 'SwimJSCC' should be 'SwinJSCC' to match reference [11].","section":"Table I"},{"comment":"The description of the loss functions says MAE with GAN is used for the proposed framework, but no training details, model weights, or code are provided, making it hard to replicate the reported curves.","section":"Section IV.A"},{"comment":"The phrase 'without the additional transmission costs' in the description of Eq. (4) conflicts with the later sparse-coding discussion; the paper should consistently distinguish between the image-only overhead and the digital-signal overhead.","section":"Section III.A"}],"recommendation":"reject","confidential_remarks":"The paper has an interesting kernel—using MAE mask positions as a side channel for digital data—and the overhead tables suggest the authors have run some experiments. However, the core digital-data path as written is not a functioning system: the bit-to-mask mapping is absent, Algorithms 1 and 2 invert incorrectly in the high-mask-ratio regime, and the BER evaluations are detached from any concrete modulation or channel-coding scheme. These are not cosmetic flaws; they invalidate the central claims as stated. A resubmission would need to specify a bijective and length-prefixed mapping, fix the sparse-matrix invertibility with an explicit flag or known-majority convention, and describe the physical transmission and error model for the index list before the reported BER can be trusted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a real idea with load-bearing holes. The novelty—using MAE mask locations as a side channel for digital bits in SemCom—is new relative to the cited literature, and the sparse-encoding module to compress mask tokens is a reasonable attempt to reduce overhead. It deserves credit for reusing a pre-trained MAE without redesigning the codec, and the simulations do show the image path behaving sensibly under AWGN and Rayleigh channels.\n\nBut the paper as written does not support the headline claim that digital signals can be reliably transmitted without significant overhead. Three soft spots, in increasing order of severity. First, the overhead analysis in Sec. III-C is internally inconsistent: the abstract says \"without significant communication overheads,\" while Sec. III-B admits mask-token transmission is a significant cost, and the formula (N - Lb)*Le + Lm is presented as decreasing with Lb without modeling how Lm grows with the bit string. Second, and load-bearing, the bit-to-mask mapping is never specified. Section III-A only says masked positions are \"determined by the indices of '1's,\" with no padding rule, no length prefix, no combinatorial code. Without that, the reported BER is not tied to any concrete transmission. Third, the pseudocode in Algorithms 1 and 2 is not invertible as written. Algorithm 1 stores the minority-class indices (either 0s or 1s), but Algorithm 2 unconditionally sets mask[:, sparse_matrix] = 1. If the minority class was the zeros, the receiver reconstructs the complement of the true mask—a bitwise inversion of the digital signal. There are also undefined variables (`ids_restore`) and an off-by-one in the length recovery, so the algorithm cannot be implemented from the text. No channel-coding or error protection for the index list is described, and since a single index error propagates through the argsort, the BER curves in Fig. 4 are not reproducible.\n\nThe reader's REJECT verdict is fair. The idea is salvageable—specify an invertible binary-to-mask code, define the index transmission scheme, and either fix the overhead accounting or drop the \"no overhead\" wording. But as is, the central claim is unsupported.\n\nI'd send this to a serious referee if I were an editor, because the core concept is novel enough to warrant expert guidance, but I wouldn't accept it in this form. For a reading group, it's a good case study in how a neat hack can hide a missing system. I wouldn't cite it yet.\n\nRecommendation: reject in current form, invite resubmission after the mapping and channel scheme are made concrete and the algorithms actually invert.","headline":"A novel idea for digital-signal SemCom, but the missing bit-to-mask mapping and non-invertible algorithms undermine the central claim.","tokens_in":15751,"tokens_out":2180,"would_cite":false,"duration_ms":21798,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a semantic-communication scheme that encodes a digital bit stream into the mask locations of an image and uses a pre-trained Masked Autoencoder to recover both the image and the bits, with a sparse encoding module to…","keywords":["semantic communication","digital signal transmission","masked autoencoder","carrier image","sparse encoding","mask tokens","joint source-channel coding","bit error rate"],"falsifier":"Take a fixed image and a binary message, map it to mask positions by the paper's unspecified rule, transmit the sparse matrix over the exact AWGN channel used in the simulations with no extra error protection, and count how often any index in the sparse matrix is corrupted; the scheme's bit error rate cannot fall below the index error rate, so a measurement showing index errors far larger than the simulated BER would falsify the claim that the digital data survives at the reported reliability.","tokens_in":14705,"feed_emoji":"🖼️","tokens_out":4836,"duration_ms":47778,"temperature":0.7,"pith_summary":"This paper tries to show that digital signals, which carry no semantic features, can still be transmitted through a semantic communication system by using an image as a carrier. The idea is to convert the binary data into a pattern of masked patches on the image and let a pre-trained Masked Autoencoder (MAE) reconstruct both the image and, from the mask locations, the original bits. A sparse encoding module is added at the transmitter so the mask information is sent as a sparse matrix rather than fixed-length mask tokens, cutting transmission overhead. Simulations over AWGN and Rayleigh fading channels are used to argue that the approach keeps image quality and bit error rate acceptable even at high mask ratios, and that its transmission cost falls below existing image-only semantic communication schemes when the same binary data is added.","feed_headline":"Digital signals ride on masked images with no extra overhead","feed_subtitle":"A pre-trained masked autoencoder reconstructs both the picture and the hidden bits, cutting transmission overhead.","key_machinery":"The binary-to-mask mapping is the carrier mechanism: a message of $L_b$ bits selects which of the $N$ image patches are masked, and the set of masked indices is transmitted as a sparse matrix along with the latent representation of the unmasked patches. The Masked Autoencoder (MAE) — an asymmetric encoder-decoder trained to fill in masked patches from visible ones — provides the reconstruction engine, since it is robust to where the masks sit. The sparse encoding module chooses the smaller of the set of masked or unmasked indices to transmit, and a masking-plus-argsort routine at the receiver rebuilds the full mask token sequence from that sparse set.","core_discovery":"The central claim is that a binary stream can be mapped to the locations of masked patches in an image, and that a pre-trained Masked Autoencoder used as a semantic encoder/decoder can recover both the original image and the digital bits from the unmasked patches' latent representations plus a sparse encoding of the mask indices. Because the MAE was designed to reconstruct images from visible patches alone, the approach needs no new joint source-channel coding structure for the digital data; the bits ride along in the mask pattern. The paper further claims that encoding the mask tokens as a sparse matrix, which is restored at the receiver, improves bit error rate over sending mask tokens directly and reduces the number of transmitted bits, with total overhead decreasing as the binary message grows longer.","pith_inferences":["The paper does not state this, but the digital capacity is bounded by the number of distinct mask patterns the image admits, roughly $\\log_2 \\binom{N}{k}$ bits for $k$ masked patches, which sets an upper limit that could be tested against the reported overhead numbers.","The scheme is effectively a steganographic side channel: the digital payload is recoverable by anyone who knows the patch indexing and mask convention, so it may serve covert or watermarking applications, though the paper does not discuss security.","A natural extension the paper leaves implicit is to add error-correcting codes over the index set or to learn the mask generation rather than using a fixed binary mapping; either could improve the bit error rate at low SNR without changing the MAE backbone."],"forward_implications":["If the approach works as claimed, any pre-trained MAE can serve as a semantic codec for digital data without retraining, so digital signals can be added to existing image-oriented semantic communication links at near-zero extra cost.","Longer binary messages imply higher mask ratios, and within the MAE's 10 to 80 percent mask-ratio range the transmitted bits decrease as the message grows, making the overhead curve decrease with payload size.","The sparse-matrix encoding is claimed to give lower bit error rate than transmitting mask tokens directly, implying that index-set sparsity itself is a useful source of error resilience.","Because the mask pattern carries the data, the image and the digital message share the same physical transmission, so a single channel use delivers both, which no current image-only semantic communication model does."],"supporting_citations":[{"why":"Supplies the pre-trained Masked Autoencoder and the evidence that it reconstructs images across a wide range of mask ratios, which the whole framework leans on.","marker":"[9]"},{"why":"Supplies the MAE with GAN loss variant used as the framework's reconstruction backbone and benchmark.","marker":"[12]"},{"why":"Provides a deep image-transmission baseline whose transmission-bit count the proposed approach is compared against in the overhead table.","marker":"[10]"},{"why":"Provides a transformer-based image-transmission baseline also used for the overhead comparison.","marker":"[11]"}],"fun_headline_variants":["Carrier images hide digital bits in mask patterns","MAE turns image masks into a digital communication channel","Binary data rides on masked image patches, saving bandwidth","SemCom via carrier images: digital bits in mask locations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire digital-data recovery depends on the masked-patch indices being reconstructed at the receiver without error, yet the paper never specifies how the bits are actually mapped to mask positions nor what channel coding protects those indices, so the reported bit error rate is not tied to any concrete transmission scheme.","fun_headline_variants_meta":{"raw":{"variants":["Carrier images hide digital bits in mask patterns","MAE turns image masks into a digital communication channel","Binary data rides on masked image patches, saving bandwidth","SemCom via carrier images: digital bits in mask locations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000382,"raw_usage":{"total_tokens":2006,"prompt_tokens":908,"completion_tokens":1098,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":1035}},"tokens_in":524,"tokens_out":1098,"duration_ms":8416,"temperature":1.0,"reasoning_tokens":1035,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:03:53.009061+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed image and a binary message, map it to mask positions by the paper's unspecified rule, transmit the sparse matrix over the exact AWGN channel used in the simulations with no extra error protection, and count how often any index in the sparse matrix is corrupted; the scheme's bit error rate cannot fall below the index error rate, so a measurement showing index errors far larger than the simulated BER would falsify the claim that the digital data survives at the reported reliability.","supporting_citations":[{"cited_title":"Masked Autoencoders Are Scalable Vision Learners,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained Masked Autoencoder and the evidence that it reconstructs images across a wide range of mask ratios, which the whole framework leans on."},{"cited_title":"Masked Auto-Encoders Meet Generative Adversarial Networks and Beyond,","cited_arxiv_id":null,"evidence_quote":"Supplies the MAE with GAN loss variant used as the framework's reconstruction backbone and benchmark."},{"cited_title":"Predictive and Adaptive Deep Coding for Wireless Image Transmission in Semantic Communication,","cited_arxiv_id":null,"evidence_quote":"Provides a deep image-transmission baseline whose transmission-bit count the proposed approach is compared against in the overhead table."}],"review_version":1}