{"id":"65cf894f-39e9-4203-b812-c22d44abfe3a","arxiv_id":"2412.00334","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Masking 75% of image patches during client-side training and moving deep layers to the server yields faster, cheaper federated ViT training with modest accuracy gains.","lead":"EFTViT is a federated learning scheme in which phones and other small devices train only the first few layers of a vision transformer on randomly masked image patches, while a server trains the deeper layers on uploaded features. The authors report that this lowers client-side computation by up to 5.6x and training time by up to 3.1x while improving accuracy by up to 2.46% over existing efficient federated tuning methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 3.6's complexity formula undercounts client backward cost by ignoring frozen global-module layers, so the headline 5.6x computational savings is not credibly supported.","rationale":"The reader identified the masking-redundancy assumption as weakest, but the more load-bearing concern is the correctness of the efficiency numbers. Even if masking works perfectly, the central contribution is the claimed reduction in client computation; that claim is supported by a complexity analysis (Section 3.6) that incorrectly omits backward cost through frozen layers. This is a specific, checkable technical error, not a matter of dataset transfer. The table values appear to conflict with the formula, leaving the efficiency claim ambiguous. Because the paper provides no code and no measurement methodology, a reader cannot resolve this without further evidence. This direct threat to the headline efficiency claim is more decisive than the image-redundancy prior, which is at least empirically demonstrated in Figure 2. I therefore disagree with the reader's weakest-assumption ranking. The correct verdict should be REJECT until the complexity account and the reported GFLOPs are reconciled with an actual implementation.","tokens_in":14337,"tokens_out":13405,"duration_ms":119554,"concrete_test":"Reproduce the client-side GFLOPs for one EFTViT training step using a standard FLOP profiler or manual count: ViT-B/16, 224x224, n=196, d=768, rm=0.75, M=2 local trainable layers, N=10 frozen global layers, with the global module kept in the autograd graph (as required to train the local module). If the measured forward+backward GFLOPs exceed the paper's 2.997 by more than 20%, the claimed 5.6x improvement over FedRA/Fed-LoRA is overstated. Also verify whether the §3.6 formula predicts a different value than the table; a discrepancy indicates the complexity analysis is not the basis for the reported numbers.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that EFTViT reduces local training cost by up to 5.6x and time by up to 3.1x, but the paper's own complexity analysis appears internally inconsistent. In Section 3.6, the backward-propagation time is stated as O(10·(NT−N)·(1−rm)·n·d² + 4·(NT−N)²·(1−rm)²·n²·d), i.e., it counts only the M=NT−N trainable local layers. However, the client loss (Eq. 1) is computed after the frozen N-layer global module; to obtain gradients for the local-module parameters, automatic differentiation must backpropagate through all NT layers, computing activation gradients even where parameters are frozen. The backward cost is therefore proportional to NT, not NT−N. With NT=12 and N=10, the formula undercounts backward work by roughly 6x. The table's reported 2.997 GFLOPs for EFTViT is suspiciously close to one-quarter of the Fed-Full value (12.005 GFLOPs), which would be consistent with full-depth backward at 25% tokens, contradicting the §3.6 formula. This ambiguity means the reader cannot tell whether the headline efficiency numbers are computed correctly or artificially reduced by the flawed formula. Since efficiency is the paper's primary contribution, this is a load-bearing correctness concern.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EFTViT, a hierarchical federated learning framework for Vision Transformers on resource-constrained clients. Each client trains a lightweight M-layer local module and a classification head on randomly masked image patches (default rm=0.75), while a central server trains the N-layer global module on median-sampled intermediate patch features uploaded from clients. The authors claim that EFTViT reduces local training computational cost by up to 5.6x, reduces local training time by up to 3.1x, improves accuracy by up to 2.46% over existing resource-constrained FL baselines, and provides privacy benefits through masking and the median sampling strategy. Experiments on CIFAR-10, CIFAR-100, and UC Merced Land-Use under two Dirichlet heterogeneity settings are reported in Tables 1-3 and Figures 4-7.","tokens_in":14660,"tokens_out":10351,"duration_ms":86507,"significance":"If the reported results hold, EFTViT is a practically valuable recipe for federated ViT training on edge clients: it combines an image-masking prior with a hierarchical parameter split and a class-balancing upload mechanism, and it shows consistent gains over several PEFT-based baselines on three benchmarks. The paper's strengths are its clear motivation (Figure 2), the breadth of compared methods, the convergence-speed analysis, and the fact that the central mechanism is simple and easy to implement. However, the absence of multiple seeds, the lack of code, the internal inconsistency in the complexity accounting, and the unresolved abstract discrepancy mean that the headline efficiency and accuracy claims should be treated with caution.","major_comments":[{"comment":"The backward-propagation complexity in Section 3.6 is undercounted. Because the client loss in Eq. (1) is computed after the frozen N-layer global module, gradients for the trainable local-module parameters phi_k must be backpropagated through all NT layers, even though no parameter gradients are accumulated for the frozen layers. The backward term O(10*(NT-N)*(1-rm)*n*d^2 + ...) should therefore scale with NT, not NT-N; with NT=12 and N=10 this undercounts backward work by roughly a factor of 6. The reported GFLOPs in Table 3 (2.997, close to 12.005/4) appear to correspond to a forward-only pass, so the paper should state explicitly that the GFLOPs column is forward-only and either correct the complexity formula or qualify the abstract's 'computational cost' claim to match the measured TTPR.","section":"Section 3.6, Eq. (1)-(2)"},{"comment":"The listing abstract reports 'up to 28.17% accuracy improvement, reduces local training computational cost by up to 2.8x, and cuts local training time by up to 4.4x', whereas the full-text abstract and Section 1 report 'reduces the computational cost of local training by as much as 5.6x, lowers the local training time by up to 3.1x, and improves the accuracy by up to 2.46%'. These are materially different claims and must be reconciled; as presented, a reader cannot tell which numbers are the actual results.","section":"Abstract and Section 1"},{"comment":"All experimental numbers appear to come from single runs with no error bars or significance tests. Several decisive comparisons are very small: on CIFAR-10 with beta=0.1, EFTViT (98.12) is only 0.13 points above Fed-Prompt and FEDBFPT (97.99), and on CIFAR-100 with beta=1.0 the margin over Fed-Bias is 0.14 points. In addition, the key hyperparameters rm, M, and the sampling threshold are selected by ablations on the same test benchmarks (Figures 6-7, Tables 4-5), which makes the reported accuracies optimistic. The 'state-of-the-art' and 'consistently outperforms' claims need multi-seed results to be convincing.","section":"Section 4.3, Tables 1 and 3"},{"comment":"The privacy evaluation is qualitative only: it shows reconstructed images and states that they are 'severely distorted' but provides no quantitative metric (e.g., PSNR/SSIM) and no comparison against reconstruction from full features or from gradients. Also, the claim that median sampling 'erases' data distribution privacy is stronger than what is demonstrated: the strategy equalizes per-class sample counts, but other distributional information in the uploaded features is not analyzed.","section":"Section 4.3, Privacy Analysis"}],"minor_comments":[{"comment":"The loss in Eq. (1) is written as a sum over classes p(y=i) log(...) but does not index over samples or divide by the batch size; please rewrite it as an empirical expectation over the local data.","section":"Eq. (1)"},{"comment":"The statement that 'the entire images are adopted for inference on each client' while training uses only 25% of patches creates a train/inference distribution shift that should be discussed explicitly, since it affects how the reported accuracy should be interpreted.","section":"Section 3.3"},{"comment":"The caption says masking to 0.75 'minimally affects' performance, but the reported CIFAR-100 drop is 3.41 points; please qualify this as a small but non-negligible degradation.","section":"Figure 2"},{"comment":"No code or repository is provided, and the paper does not state whether the reported numbers are averaged over multiple runs; please add reproducibility details, including the number of seeds and hardware/software versions.","section":"Section 4.2"},{"comment":"The distinction between the GFLOPs column (forward propagation only, per Section 4.3) and the complexity analysis in Section 3.6 (forward plus backward) should be made explicit in the table caption and in the abstract's use of the term 'computational cost'.","section":"Table 3 and Section 3.6"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the backward-complexity formula is valid and is the main technical issue; the table's GFLOPs value strongly suggests that the reported efficiency numbers are forward-only, which the manuscript never makes clear. The abstract discrepancy between the listing and the full text is also serious and needs to be fixed. I would ask the authors for a corrected complexity analysis, multi-seed experiments, and a reproducibility statement or code release before considering this for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my take on EFTViT. The paper combines three mostly known ideas – client-side patch masking, split learning with a server-trained global module, and class-balanced median sampling – into a system that trains ViTs on constrained clients. The experiments are consistent: EFTViT beats the PEFT baselines on all three datasets, converges much faster, and the ablation on masking ratio shows 75% masking costs only a few points. The privacy reconstruction analysis is a nice addition.\n\nThe central problem is the complexity analysis in Section 3.6. The backward-propagation cost is written as proportional to NT−N (the local layers only), on the reasoning that the global module parameters are frozen. That is wrong. Backpropagating a loss computed after the global module to the local-module parameters requires computing activation gradients through every global-module layer, even if those parameters are not updated. A frozen layer still needs Jacobian-vector products for its input gradient. So the backward cost is proportional to the full depth NT, not just the trainable depth. The table itself confirms this: EFTViT's GFLOPs (2.997) are nearly exactly 25% of Fed-Full's (12.005), which is what you get by multiplying the entire forward+backward pass by the 25% token fraction. If the paper's formula were right, EFTViT should come in closer to about one-sixth of Fed-Full's backward cost, i.e., well below 2.997. So the headline 5.6x speedup over Fed-LoRA/FedRA is not being produced by the 'frozen global module' trick; it's produced by the masking ratio and the baselines' extra LoRA FLOPs. The authors should revise the formula and restate what the actual source of the savings is.\n\nOther soft spots: no code or data, no error bars (all results appear to be single-run), and key hyperparameters (rm, M, sampling threshold) are selected on the same benchmarks that are reported as the main results. The ablations do show robustness, so this is more of a reproducibility concern than a fatal flaw. Also, the abstract in the arXiv listing contradicts the full-text abstract (28.17% vs 2.46% accuracy, 2.8x vs 5.6x), which needs cleaning up if the paper is to be cited.\n\nWith those fixes, this is a reasonable systems contribution. The idea is useful and the experiments suggest it works. I would send it to peer review, but I would push back on the efficiency claims until the complexity analysis is corrected.","headline":"Useful combination of known components that works in experiments, but the complexity analysis in §3.6 is wrong; the actual savings come from masking, not from freezing the global module.","tokens_in":15166,"tokens_out":6648,"would_cite":false,"duration_ms":57179,"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":"Federated ViT training can be made 5.6x cheaper on clients by masking 75% of each image, while improving accuracy by up to 2.46% over prior resource-constrained methods.","keywords":["federated learning","vision transformer","masked image modeling","resource-constrained clients","hierarchical training","data heterogeneity","intermediate feature privacy","parameter-efficient fine-tuning"],"falsifier":"Run EFTViT with $r_m = 0.75$ on a dataset whose discriminative information is spread uniformly across the whole image (e.g., fine-grained texture or medical pathology images) and compare against Fed-Full: if the accuracy gap grows well beyond the roughly 1-2% seen on CIFAR and UC Merced, the redundancy premise is falsified. A cheaper check is to sweep the masking ratio on any new dataset and see whether accuracy collapses before 75% is reached.","tokens_in":14162,"feed_emoji":"⚡","tokens_out":9232,"duration_ms":74134,"temperature":0.7,"pith_summary":"EFTViT is a proposal for training Vision Transformers in federated settings where client devices are too weak to run the full model. The paper argues that images are redundant enough that randomly masking 75% of the patches before local training barely hurts accuracy while cutting local computation by up to 5.6x and local training time by up to 3.1x. To make this work, the model is split into lightweight local modules on clients and a larger global module on the server, with the server trained on intermediate patch features rebalanced by a median sampling strategy. If the claims hold, resource-constrained edge clients could train all ViT parameters collaboratively rather than falling back to parameter-efficient fine-tuning, and the masked patches also make uploaded features harder to reconstruct into images. The experiments report accuracy gains up to 2.46% over existing resource-constrained FL methods on CIFAR-10, CIFAR-100, and UC Merced Land-Use.","feed_headline":"Masking 75% of image patches cuts federated ViT training cost 5.6x","feed_subtitle":"Clients use only a quarter of each image and still beat prior resource-constrained FL methods by up to 2.46%.","key_machinery":"The central object is the masking-plus-hierarchy training scheme. Images are patchified into 16x16 tokens, a fraction $r_m = 0.75$ of patches is randomly dropped, and only the remaining tokens, with positional embeddings preserved, pass through the client's local module of $M = 2$ Transformer layers; the server's global module of $N = 10$ layers is trained on the uploaded patch features. A median sampling strategy balances the patch-feature dataset so each class contributes the same number of samples, hiding client data-distribution statistics. The complexity analysis ties the savings to the formula $\\mathcal{O}((15N_T - 10N)(1 - r_m) n d^2 + (6N_T - 4N)(1 - r_m)^2 n^2 d)$, where $N_T$ is the total number of Transformer layers, $n$ the number of patches, and $d$ the latent dimension, so both a larger server-side share $N$ and a higher masking ratio $r_m$ reduce client cost.","core_discovery":"The central claim is that ViT training for federated classification can be made cheap enough for resource-constrained clients by exploiting image redundancy: divide each image into non-overlapping patches, randomly discard 75% of them, and train the client-side layers only on the remaining quarter. The paper reports that this masking step reduces the client's forward-pass cost roughly fivefold with minimal accuracy loss, and that the saved budget can be reinvested in training all model parameters across the hierarchy instead of only a small tunable subset. Concretely, each client trains two Transformer layers and a classification head on masked patches, while the server trains the remaining ten Transformer layers on patch features that clients balance across classes using the median class count before upload. Across CIFAR-10, CIFAR-100, and UC Merced Land-Use with high and low data heterogeneity, EFTViT is reported to beat parameter-efficient baselines (head-only, bias, prompt, LoRA) and partial-training methods (FEDBFPT, FedRA), reaching target accuracy in fewer rounds and with 5.6x lower local computation and 3.1x lower per-round training time.","pith_inferences":["Beyond the three tested benchmarks, the 75% masking sweet spot should transfer only to datasets whose classification signal is spatially redundant; fine-grained or small-object datasets are the natural stress test.","The reconstruction-attack experiment is heuristic, so the privacy claim is best read as 'harder to reconstruct' rather than a formal guarantee; a determined attacker with a learned image prior might still recover coarse content.","The efficiency comparison is against PEFT and partial-training baselines, not against full-parameter FedAvg; reading the reported numbers against the Fed-Full row suggests the honest trade is roughly 0.3 to 2.3% accuracy for a 4x client-compute reduction.","Median sampling hides per-class counts, but it may still leak correlations between features and labels or class co-occurrence; differential privacy, which the paper notes is orthogonal, could close that gap."],"forward_implications":["Client devices can participate in full-parameter ViT training instead of being limited to small tunable subsets, which is what the accuracy gains over head, bias, prompt, and LoRA baselines are attributed to.","Uploaded 25%-patch features are harder to turn back into images, so masking acts as a built-in content-privacy layer on top of the usual intermediate-feature exchange.","Median sampling equalizes class counts before upload, so the server cannot infer client preference skew from data distribution statistics.","Convergence to 85% accuracy is claimed in as few as 3 to 6 rounds on the tested benchmarks versus 9 to 119 rounds for baselines, which translates to large reductions in total wall-clock time including communication.","The method's accuracy changes little when the local module depth $M$ is 2, 4, or 6 layers, so the same framework adapts to clients with different compute budgets."],"supporting_citations":[{"why":"Supplies the ViT-B patch-based backbone architecture that EFTViT splits into local and global modules.","marker":"[5]"},{"why":"Prior hierarchical client-server split with intermediate features; EFTViT extends this idea to ViTs and masked inputs.","marker":"[8]"},{"why":"Masked autoencoders demonstrate that images are redundant enough to learn from a small subset of patches, motivating the masking strategy.","marker":"[10]"},{"why":"LoRA is adapted as a federated parameter-efficient fine-tuning baseline that EFTViT is compared against.","marker":"[11]"},{"why":"FedAvg defines the federated learning protocol and evaluation setup used in the experiments.","marker":"[24]"},{"why":"FedRA is a state-of-the-art resource-constrained transformer FL baseline that EFTViT must beat.","marker":"[27]"},{"why":"Masked image modeling evidence that removing a large fraction of patches still permits representation learning.","marker":"[31]"},{"why":"FEDBFPT is a partial-parameter-training baseline that progressively optimizes shallow layers, serving as a direct comparison point.","marker":"[32]"},{"why":"BitFit bias-tuning is adapted as a parameter-efficient federated baseline for comparison.","marker":"[36]"}],"fun_headline_variants":["Mask 75% of image patches for 5.6x cheaper federated ViT training","Toss 75% of patches, cut federated ViT training cost by 5.6x","EFTViT: 75% patch masking makes federated ViT training feasible on edge clients","Train federated ViTs on a quarter of each image, get 5.6x speedup","Mask 75% of patches; federated ViT training costs 5.6x less"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that images carry enough redundancy that randomly discarding 75% of the patches before training barely hurts classification accuracy, and if that fails on a new dataset, at a different patch size, or under stronger data heterogeneity, the efficiency gains will come at an unacceptable accuracy cost.","fun_headline_variants_meta":{"raw":{"variants":["Mask 75% of image patches for 5.6x cheaper federated ViT training","Toss 75% of patches, cut federated ViT training cost by 5.6x","EFTViT: 75% patch masking makes federated ViT training feasible on edge clients","Train federated ViTs on a quarter of each image, get 5.6x speedup","Mask 75% of patches; federated ViT training costs 5.6x less"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3551,"prompt_tokens":1071,"completion_tokens":2480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":2355}},"tokens_in":687,"tokens_out":2480,"duration_ms":14168,"temperature":1.0,"reasoning_tokens":2355,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:28:48.706942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EFTViT with $r_m = 0.75$ on a dataset whose discriminative information is spread uniformly across the whole image (e.g., fine-grained texture or medical pathology images) and compare against Fed-Full: if the accuracy gap grows well beyond the roughly 1-2% seen on CIFAR and UC Merced, the redundancy premise is falsified. A cheaper check is to sweep the masking ratio on any new dataset and see whether accuracy collapses before 75% is reached.","supporting_citations":[{"cited_title":"Group knowledge transfer: Federated learning of large cnns at the edge","cited_arxiv_id":null,"evidence_quote":"Prior hierarchical client-server split with intermediate features; EFTViT extends this idea to ViTs and masked inputs."},{"cited_title":"Communication- efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"FedAvg defines the federated learning protocol and evaluation setup used in the experiments."},{"cited_title":"Fedra: A ran- dom allocation strategy for federated tuning to unleash the power of heterogeneous clients","cited_arxiv_id":null,"evidence_quote":"FedRA is a state-of-the-art resource-constrained transformer FL baseline that EFTViT must beat."},{"cited_title":"Revealing the dark secrets of masked im- age modeling","cited_arxiv_id":null,"evidence_quote":"Masked image modeling evidence that removing a large fraction of patches still permits representation learning."},{"cited_title":"Fedbfpt: An efficient federated learning framework for bert further pre-training","cited_arxiv_id":null,"evidence_quote":"FEDBFPT is a partial-parameter-training baseline that progressively optimizes shallow layers, serving as a direct comparison point."}],"review_version":1}