{"id":"89ae442e-42c4-4c0c-8e0f-3daac3b4af4d","arxiv_id":"2504.14188","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"FedC4 combines graph condensation with selective client-to-client sharing, reporting better accuracy and up to 1000x lower communication in federated graph learning, though its privacy proof has a gap.","lead":"FedC4 is a new framework for federated graph learning in which clients exchange tiny synthetic graphs instead of raw node data, cutting communication costs and reducing node-level exposure. A smart generalist should read this because practical federated learning for graph data is a key privacy bottleneck in cross-institution settings such as finance and healthcare.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The privacy proof's key step (Eq. 21) is undrived: Eq. 16 does not depend on the synthetic graph, and with FedC4's fixed condensation ratios the claimed O(m/n) vanishing does not hold.","rationale":"The reader's weakest assumption identifies the same load-bearing defect: the privacy proof's transition from a gradient-difference bound to a synthetic-embedding sensitivity bound is unsupported, and the use of fixed condensation ratios invalidates the claimed O(1) shrinking. My closer reading confirms this and sharpens it: the loss in Eq. (16) is defined on the original graph only, so the entire proof may be about a quantity that does not involve S; the Hessian term in Eq. (21) is not defined; and the m/n bound is not derived. This matters because privacy protection is one of the three headline contributions and is repeated in the abstract, the introduction, Section 4, and Section 5.5. The empirical evaluation is broad, and the released code is a positive sign, but a central theoretical claim that is internally incomplete cannot be accepted as evidence for the strongest claim. The reader's REJECT stays unchanged; a major revision would need either a correct sensitivity analysis for the actual gradient-matching objective or an honest restatement of the privacy contribution as an empirical robustness property rather than a proof.","tokens_in":16441,"tokens_out":6957,"duration_ms":62110,"concrete_test":"Independently re-derive Eq. (21)-(22) from the actual condensation objective, Eq. (6), by writing the condensed embedding H_syn as the output of the gradient-matching optimization and then differentiating with respect to removal of an original node j. If the derivation requires an unstated assumption, such as uniform invertibility of the Hessian of the gradient-matching loss with respect to the synthetic graph parameters, or that optimization returns the same local optimum with and without node j, state it explicitly and check it against the GCond-style synthesis used in FedC4. If Eq. (21) cannot be derived from Eq. (6) without such assumptions, the privacy theorem is not established by the argument in Sec. 4.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sec. 4.1 does not connect the leave-one-out gradient bound to the synthetic graph whose privacy it claims to establish. Eq. (16) defines L(S, G) as a loss on the original graph G; the synthetic graph S appears only in the notation. Removing a node changes that loss gradient by O(1/n), a standard finite-sample property that by itself says nothing about how the condensed graph S changes. Eq. (21) then asserts Δ_j = ||H_syn − H_syn^{-j}|| ≤ ||H^{-1}|| · ||Δ∇||, introducing an undefined Hessian and skipping the condensation optimization: Eq. (6) minimizes a gradient-matching objective between original and synthetic losses, but the proof never differentiates the synthetic output with respect to a removed original node through that optimization. The factor m in Eq. (22) appears without derivation; even if an influence-function expansion were intended, one would need H to be uniformly invertible and the condensed optimum to move smoothly with G. Moreover, FedC4 uses fixed condensation ratios (8%, 4%, 2%, ...), so m/n is constant in the reported experiments; the conclusion that influence 'diminishes as the size of the original graph increases' does not follow for the settings actually evaluated. The paper's own caveat that the proof applies only to non-overlapping subgraphs narrows the result, but it does not repair the missing derivation. Because 'intrinsic privacy-preserving nature' is presented as a validated theoretical contribution, this unsupported step is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FedC4, a client-to-client federated graph learning framework in which each client condenses its local graph into a small synthetic graph and then selectively shares statistics, representative node embeddings, and reconstructed graph topology with other clients through three modules: Customizer (CM), Node Selector (NS), and Graph Rebuilder (GR). The authors categorize prior FGL work into server-client and client-client paradigms, argue that fine-grained personalized client-client communication is preferable, and claim that graph condensation reduces communication cost and provides intrinsic privacy protection. They report experiments on eight graph datasets showing average accuracy improvements over baselines and large communication/training cost reductions, together with ablations and hyperparameter studies.","tokens_in":16662,"tokens_out":7964,"duration_ms":69860,"significance":"If the empirical claims hold, FedC4 is a potentially useful engineering contribution: it integrates graph condensation with selective client-client communication, releases code, and evaluates on a broad set of transductive, inductive, heterogeneous, and large-scale benchmarks. The proposed categorization of FGL into S-C and C-C paradigms with levels of personalization is a helpful organizing viewpoint. However, the paper's second central claim—that graph condensation is intrinsically privacy-preserving and that this is proven in Sec. 4.1—is not supported. The privacy proof is invalid as written, and the experimental privacy study in Sec. 5.5 does not test privacy. Because the privacy guarantee is advertised in the abstract, conclusion, and Q4, this is a load-bearing weakness rather than a presentation issue.","major_comments":[{"comment":"The privacy proof does not connect the original-graph gradient bound to the change in the synthetic graph. L(S,G) in Eq. (16) depends only on the original graph; the synthetic graph S appears only as a placeholder, so the leave-one-out gradient bound in Eqs. (18)-(20) is a standard finite-sample property of the original loss and does not bound how the condensed graph changes when an original node is removed. Eq. (21) then asserts Δ_j = ||H_syn − H_syn^{-j}|| ≤ ||H^{-1}||·||Δ∇|| with an undefined Hessian H and no influence-function argument that differentiates the condensation optimization in Eq. (6); the factor m in Eq. (22) appears without derivation. Even if such an expansion were intended, it would require uniform invertibility of H and smooth dependence of the condensed optimum on G, neither of which is stated or shown. Finally, because the experiments fix condensation ratios (8%, 4%, 2%, 0.4%, etc.), m/n is constant in the tested regime, so the conclusion that the influence of a node 'diminishes as the size of the original graph increases' does not follow from the settings actually evaluated. The caveat limiting the proof to non-overlapping subgraphs narrows the claim but does not repair the missing derivation.","section":"Sec. 4.1, Eqs. (16)-(22)"},{"comment":"The privacy evaluation does not measure privacy. Injecting Laplace noise into the condensation process and reporting that accuracy degrades slowly is a robustness-to-noise experiment, not a test of membership inference, attribute inference, or reconstruction resistance, and no formal privacy accounting links the noise scale to a guarantee. The reference to 'Privacy for Free' [2] concerns image dataset condensation and does not automatically transfer to graph embeddings, particularly because the extension proof in Sec. 4.1 is incomplete. The claim in the conclusion that GC 'mitigates privacy risks' is therefore not supported by the experimental section as written.","section":"Sec. 5.5, Fig. 7(b)"},{"comment":"The communication cost O(C log C N'd) for FedC4 is not derived. The manuscript describes selective broadcasting to same-cluster clients (Eq. (11), Algorithm 2), but it does not specify a clustering or communication hierarchy that would introduce a log factor; in the worst case, when all clients fall into one cluster, the cost is O(C^2 N'd). The table should either state the structural assumptions under which O(C log C N'd) holds or present the worst-case cost.","section":"Sec. 4.2, Table 2"}],"minor_comments":[{"comment":"The baseline description names SGDD as one of the three FL+GC methods, but Table 1 lists a column labeled 'DosCond' and no reference for DosCond is provided; please align the text, table, and reference list.","section":"Sec. 5.1 and Table 1"},{"comment":"The Sliced Wasserstein Distance formula uses the same symbol θ both as the integration variable over the sphere and inside the supremum; as written the outer integral is redundant and the expression is not the standard SWD. Please rewrite with distinct variables or cite the precise definition used in the implementation.","section":"Eq. (12)"},{"comment":"References [4] and [5] are identical (both point to the same ICDE 2024 paper), while the text uses them for different purposes (MCond and self-expressive reconstruction); please correct the citations.","section":"References [4] and [5]"},{"comment":"The text says Fig. 5(a) shows the impact of varying α and β and Fig. 5(b) shows the effect of τ, but the figure panels are labeled 'Node Selector' and 'Graph Rebuilder' and appear to show the opposite; please align the text with the figure.","section":"Sec. 5.4 and Fig. 5"},{"comment":"The clustering threshold δ in Algorithm 2 is not specified or studied in the hyperparameter analysis; please report its value and sensitivity.","section":"Algorithm 2"},{"comment":"The claimed 'efficiency improvement up to 1000x' is not quantified in the main text; Fig. 6 shows accuracy-versus-time curves, but the basis for the 1000x factor should be made explicit.","section":"Abstract and Sec. 1"}],"recommendation":"reject","confidential_remarks":"The empirical comparison is extensive and the code is released, but the theoretical privacy contribution is a central advertised result of the paper. The current proof cannot be repaired by local edits: the key transition from an original-graph gradient bound to a synthetic-embedding sensitivity bound is absent, and the fixed-ratio experiments contradict the claimed O(m/n) vanishing rate. A valid privacy analysis or a substantial reframing of the paper's claims would be required, so I recommend rejection rather than minor revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about FedC4. The empirical work is substantial: it actually builds graph condensation into client-to-client federated graph learning, with three modules (customizer, node selector, graph rebuilder) that turn a broadcast-heavy C-C design into selective, personalized exchanges. Eight datasets, a dozen baselines, ablations, hyperparameter sweeps, and code. That is real engineering, and the accuracy and communication gains are plausible. The second thing is that the theory does not hold. The privacy proof in Sec 4.1 is a non-sequitur: Eq. (16) is a loss on the original graph, removing a node gives the standard O(1/n) gradient-difference bound, and then Eq. (21) jumps to a bound on synthetic-embedding differences through an undefined Hessian. Nothing in the proof differentiates the synthesized graph with respect to the original node removal through the condensation optimization. The 'm' factor in Eq. (22) appears from nowhere, and because the experiments use fixed condensation ratios, m/n is constant, so the claim that influence 'diminishes as the size of the original graph increases' does not describe the evaluated settings. The paper's own caveat about non-overlapping subgraphs narrows the claim further. This is load-bearing because privacy is a headline contribution. The communication complexity O(C log C) is also asserted, not derived; nothing about SWD clustering guarantees log C groups, and the server-side aggregation cost is omitted. Table 1 is messy: for some rows baselines are missing, and the 1.73% average gain is not transparently computed, though the overall win pattern looks real. Where does that leave us? The core idea—condense locally, exchange synthetic graphs selectively, rebuild graphs at the receiver—is worth having in the literature, and the empirical evaluation is better than most FGL papers. But the theoretical contributions as written are not established, and the authors oversell them. A serious referee would send this back for major revision: redo or weaken the privacy analysis (synthetic embeddings likely reduce exposure in practice, but that is a claim about attack surface, not a proven bound), derive or soften the communication complexity, and clean up the tables. The system itself is plausible and reproducible enough to engage with. So: worth a serious referee, worth reading if you work on FGL, but do not trust the proof.","headline":"FedC4 has real empirical substance integrating graph condensation into client-to-client FGL, but the privacy proof in Sec 4.1 is a non-sequitur and the communication-cost analysis is under-derived; it deserves a serious referee but not acceptance as-is.","tokens_in":17285,"tokens_out":2182,"would_cite":true,"duration_ms":20240,"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":"FedC4 sends each client a small distilled graph instead of raw node embeddings, and reports better accuracy and up to 1000x lower communication than prior federated graph learning methods.","keywords":["federated graph learning","graph condensation","client-client communication","graph neural networks","personalization","communication efficiency","privacy preservation"],"falsifier":"Measure the maximum change in the transmitted condensed embeddings when one node is removed from a client graph of size $n$ while keeping the condensation ratio fixed (e.g., 2%); if that change does not decrease as $n$ grows, or if a membership-inference classifier can detect which nodes were used to build the condensed graph above chance, the paper's privacy conclusion would be refuted.","tokens_in":16149,"feed_emoji":"🕸️","tokens_out":6727,"duration_ms":54938,"temperature":0.7,"pith_summary":"This paper proposes FedC4, a client-to-client federated graph learning method that replaces the usual exchange of raw node embeddings with graph condensation: each client compresses its private graph into a few synthetic nodes and sends those. The authors argue that this shifts federated graph collaboration from broadcast of identical messages to fine-grained, per-client tailored communication, which they categorize as a new level of personalization. They report that FedC4 beats twelve baselines on eight datasets, with an average 1.73% accuracy gain over the strongest competitor and communication savings up to 1000x. The intended payoff is a practical regime where distributed graph learning is both more accurate and cheaper to run, with privacy improved because only condensed, not raw, node information crosses the network.","feed_headline":"Tiny synthetic graphs beat raw node sharing in federated learning","feed_subtitle":"FedC4 shaves communication up to 1,000x while improving accuracy an average 1.73 percent across eight graph datasets.","key_machinery":"The load-bearing mechanism is graph condensation used as a communication format: a client's graph $G=(A,X,Y)$ is compressed into a small synthetic graph $S=(A',X',Y')$ by gradient matching, so knowledge travels between clients as synthetic node embeddings rather than raw nodes. On top of this, three modules carry the personalization: the Customizer normalizes and selectively broadcasts embedding statistics to same-cluster clients; the Node Selector groups clients by Sliced Wasserstein distance and picks representative nodes by embedding-prototype similarity; and the Graph Rebuilder reconstructs an adjacency matrix for selected nodes through sparsified self-expressive reconstruction.","core_discovery":"The paper's central claim is that the bottleneck in client-client federated graph learning is message granularity, not aggregation. FedC4 replaces raw node-level exchanges with locally condensed synthetic graphs and makes every message client-specific through three modules, turning broadcast-style client-client communication into fine-grained personalized communication. The authors report that on eight datasets spanning transductive, inductive, heterogeneous, and large-scale settings, FedC4 outperforms twelve baselines while cutting communication cost by up to three orders of magnitude.","pith_inferences":["A sharper empirical test of the privacy claim would be a membership-inference attack on the condensed embeddings at increasing original-graph sizes; because the condensation ratios are fixed (e.g., 8%, 4%, 2%), the $O(m/n)$ bound may be constant rather than diminishing, and the paper's Laplace-noise robustness curves do not directly measure membership leakage.","The Customizer/Node Selector communication template, clustering clients by distribution distance and then sending tailored statistics rather than broadcasts, could transfer to non-graph federated learning of embeddings or features.","Since the Graph Rebuilder visibly repairs condensation-induced structural damage (lower KL divergence and density, higher homophily), the same reconstruction loss could improve centralized graph condensation even without federation."],"forward_implications":["Communication cost drops from $O(C^2 N d)$ for broadcast client-client exchange to $O(C \\log C\\, N' d)$ with $N' \\ll N$, because only condensed embeddings and statistics cross the network.","Fine-grained client-client collaboration outperforms both server-client and broadcast client-client methods across transductive, inductive, heterogeneous, and large-scale settings, with an average 1.73% gain over the strongest baseline.","Condensation ratios as small as 0.1% on large graphs such as Arxiv, Products, and Reddit can match or beat baselines that otherwise run out of memory or time.","Because only synthetic embeddings are transmitted and single-node influence is argued to be bounded, privacy exposure moves from original nodes to condensed nodes, and the method tolerates added Laplace noise without collapsing accuracy.","The framework keeps accuracy stable as the number of clients grows from 5 to 15, indicating robustness to data-partition granularity."],"supporting_citations":[{"why":"Supplies the local graph-condensation procedure (gradient matching, MLP-generated adjacency, sparsification) that FedC4 runs on each client.","marker":"[8]"},{"why":"Provides the prior proof that dataset condensation is inherently private, which the paper extends to graphs in its privacy analysis.","marker":"[2]"},{"why":"Gives the self-expressive graph reconstruction concept used by the Graph Rebuilder module.","marker":"[5]"},{"why":"The FedSage+ client-client baseline that broadcasts raw node-level information, the main approach FedC4 replaces.","marker":"[32]"},{"why":"The FedGCN client-client baseline that exchanges compressed intermediate features, used as a comparison in communication and privacy.","marker":"[29]"},{"why":"The FedDEP client-client baseline with dependency-aware communication and differential privacy, used as a comparison.","marker":"[31]"},{"why":"The GCN architecture used as the local model in every client.","marker":"[9]"},{"why":"The FedGTA server-client baseline representing the paradigm FedC4 is positioned against.","marker":"[10]"}],"fun_headline_variants":["FedC4 condenses graphs for 1000x cheaper FL","Tailored synthetic graphs slash federated learning costs","Graph condensation makes federated learning 1000x leaner","Client-specific synthetic nodes outperform raw sharing in FL","FedC4 shrinks graph messages with personalized condensation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The privacy proof assumes that the per-sample gradient bound $O(1/n)$ propagates through an inverse-Hessian factor to the synthetic embeddings, and assumes client subgraphs are disjoint; with fixed condensation ratios the resulting $O(m/n)$ bound is constant, not shrinking as graphs grow.","fun_headline_variants_meta":{"raw":{"variants":["FedC4 condenses graphs for 1000x cheaper FL","Tailored synthetic graphs slash federated learning costs","Graph condensation makes federated learning 1000x leaner","Client-specific synthetic nodes outperform raw sharing in FL","FedC4 shrinks graph messages with personalized condensation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00102,"raw_usage":{"total_tokens":4267,"prompt_tokens":873,"completion_tokens":3394,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":3315}},"tokens_in":489,"tokens_out":3394,"duration_ms":21518,"temperature":1.0,"reasoning_tokens":3315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:54:39.052444+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the maximum change in the transmitted condensed embeddings when one node is removed from a client graph of size $n$ while keeping the condensation ratio fixed (e.g., 2%); if that change does not decrease as $n$ grows, or if a membership-inference classifier can detect which nodes were used to build the condensed graph above chance, the paper's privacy conclusion would be refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The FedSage+ client-client baseline that broadcasts raw node-level information, the main approach FedC4 replaces."}],"review_version":1}