HeLo: Distinct Systems in Crypto, EDL & Federated Learning

• HeLo is a naming coincidence for three independent frameworks addressing IoT security, emotion distribution learning, and federated LoRA fine-tuning.
• The HELO cryptographic system employs ECC/ECDH, ChaCha20, and multithreading to achieve lightweight, efficient peer-to-peer security for resource-constrained IoT devices.
• HeLo-based frameworks in affective computing and federated learning leverage heterogeneous modality fusion and capacity-aware layer allocation to optimize performance under diverse constraints.

Searching arXiv for the provided HeLo/HELO variants and exact IDs to ground the article. In current arXiv usage, HeLo or HELO does not denote a single technical construct. It appears as the name of three unrelated research systems: HELO (Hybrid Encryption Lightweight Optimization), a lightweight cryptographic system for peer-to-peer IoT data transmission; HeLo, a multi-modal emotion distribution learning framework with explicit label-correlation modeling; and Fed-HeLLo, a federated foundation-model fine-tuning framework based on heterogeneous LoRA allocation. The shared label therefore identifies a naming coincidence across security, affective computing, and federated adaptation rather than a unified research lineage (Ahmed et al., 5 May 2026, Zheng et al., 9 Jul 2025, Zhang et al., 13 Jun 2025).

1. Nomenclature and domain-specific meanings

The three uses of the name differ in expansion, problem setting, and methodological core.

Variant	Expansion	Domain
HELO	Hybrid Encryption Lightweight Optimization	P2P IoT cryptography
HeLo	Heterogeneous Multi-Modal Fusion with Label Correlation	Emotion distribution learning
Fed-HeLLo	Efficient Federated Foundation Model Fine-Tuning with Heterogeneous LoRA Allocation	Federated LoRA fine-tuning

This terminological divergence matters because the internal meanings of the name are domain-specific. In the cryptographic setting, “hybrid” refers to the combination of ECC/ECDH, ChaCha20, Poly1305, ECDSA, SHA3-256, tokenization, chunking, and multithreading. In the affective-computing setting, “heterogeneous” refers to physiological and behavioral modalities and to the explicit exploitation of label correlation in emotion distributions. In the federated-learning setting, “HeLLo” refers to Heterogeneous LoRA Allocation, namely the assignment of different trainable LoRA-layer subsets to clients with different resource budgets (Ahmed et al., 5 May 2026, Zheng et al., 9 Jul 2025, Zhang et al., 13 Jun 2025).

A common misconception is to treat these works as variants of a single method family. The papers instead define independent frameworks with unrelated objectives, datasets, architectures, and evaluation criteria.

2. HELO as a lightweight cryptographic system for IoT peer-to-peer transmission

In "HELO Cryptography: A Lightweight Cryptographic System for Enhancing IoT Security in P2P Data Transmission" (Ahmed et al., 5 May 2026), HELO is a complete cryptographic system designed for end-to-end security in peer-to-peer transmission between resource-constrained IoT devices. It specifies how peers establish shared keys using ECDH over SECP256R1, derive a 256-bit symmetric key with HKDF using SHA3-256, encrypt data with ChaCha20, authenticate ciphertext with Poly1305, sign plaintext with ECDSA, and manage transfer through chunking, nonces, tokenization, and multithreading.

The design targets confidentiality, integrity, authenticity, non-repudiation, replay protection, and availability. Confidentiality is provided by ChaCha20 under a symmetric key $K_d$ derived from the ECDH shared secret $S_k$. Integrity is enforced through Poly1305 and SHA3-256 hashing; authenticity and non-repudiation are tied to ECDSA signatures bound to the sender’s private key. Replay protection is based on a per-session 128-bit nonce token $T_n$, generated via CSPRNG and treated as single-use. Availability is addressed through chunking, independent retransmission of corrupted chunks, and multithreading.

The ciphertext structure is explicitly length-preserving apart from fixed metadata. If the plaintext size is $M$ bytes, the encrypted file contains a 16-byte nonce, $M$ bytes of ChaCha20 ciphertext, and a 16-byte Poly1305 tag, for total size

$$\text{Total size} = 16 + M + 16 = M + 32 \text{ bytes}.$$

This constant 32-byte overhead is independent of message size. The protocol signs plaintext before encryption, stores the signature separately as a .sig file, verifies the MAC before decryption, and aborts if tag or signature verification fails.

The system is called lightweight for several reasons stated in the paper. It uses ECC rather than RSA, with the comparison that 256-bit ECC security ~ 3072-bit RSA; it uses ChaCha20 because of software efficiency on CPUs without AES acceleration and resistance to cache-timing attacks arising from table-based AES; and it processes data in developer-chosen chunks $C_s$ to reduce RAM usage. The reported RAM footprint for text and CSV test cases is typically about 4.7–5 MB, lower than AES/Blowfish/Fernet at about 6–6.5 MB. On TXT files, HELO had average CPU time about 0.016 s per file, lower than AES at 0.032 s, Fernet at 0.041 s, and Blowfish at 0.068 s. On CSV files up to 124 MB, HELO is reported to outperform the comparison methods; for a 124 MB CSV example, HELO is about 0.39 s versus AES about 1.50 s, Blowfish about 0.88 s, and Fernet about 1.30 s. For energy, using the model $E = P \times t$ with CPU TDP $P = 95\,\text{W}$, the paper reports substantially lower energy on TXT and CSV, including the example 124 MB, HELO ~37 J vs AES ~142.5 J. The main performance boundary is large image data: HELO is fastest on small images, but for 20–40 MB images, AES and sometimes Blowfish or Fernet outperform it in runtime and energy (Ahmed et al., 5 May 2026).

The paper’s security analysis is informal rather than game-based. It assumes an adversary who controls the communication channel and can observe, modify, inject, and replay messages, but does not compromise endpoint private keys or break standard hardness assumptions. The analysis maps component primitives to attack classes including MITM, eavesdropping, spoofing, replay, brute force, preimage and birthday attacks, tampering, and cache timing. The avalanche-effect measurements are reported at approximately 50.0–52.0% for TXT, CSV, and image files, which the paper interprets as good diffusion within realistic bounds (Ahmed et al., 5 May 2026).

3. HeLo as heterogeneous multi-modal fusion with label correlation for emotion distribution learning

In "HeLo: Heterogeneous Multi-Modal Fusion with Label Correlation for Emotion Distribution Learning" (Zheng et al., 9 Jul 2025), HeLo is a framework for emotion distribution learning (EDL) rather than single-label or binary multi-label recognition. The target is an emotion distribution

$$L = (d_1, d_2, \ldots, d_l), \quad d_k \ge 0,\ \sum_{k=1}^l d_k = 1,$$

where each $S_k$0 denotes the intensity of the $S_k$1-th basic emotion. The framework is motivated by the presence of mixed emotions and by the heterogeneity between physiological and behavioral modalities.

The model has two principal stages. The first is a modality fusion stage. Within physiological data, it uses multi-head cross-attention in the CAPF module, with EEG as the query modality and other physiological signals as keys and values. For a modality $S_k$2, the paper defines

$S_k$3

followed by

$S_k$4

The resulting physiological representation is then fused with behavioral features through an optimal transport-based heterogeneity mining module (OTHM). With physiological representation $S_k$5 and behavioral representation $S_k$6, the Wasserstein alignment objective is

$S_k$7

where $S_k$8 is the transport plan. The learned transport plan is then used to re-weight physiological tokens before a transformer encoder and concatenation with behavioral features.

The second stage is a label correlation learning stage. HeLo introduces a learnable label embedding matrix

$S_k$9

constructs a learned label-correlation matrix

$T_n$0

and aligns it with a ground-truth correlation matrix

$T_n$1

through

$T_n$2

These learned label correlations are injected directly into the Label Correlation-Driven Cross-Attention (LCDCA) mechanism: $T_n$3 A three-layer MLP and softmax produce the predicted emotion distribution $T_n$4. Training uses

$T_n$5

The framework is evaluated on DMER and WESAD. DMER uses EEG, GSR, PPG, and facial video; WESAD uses ECG, EMG, EDA, and ACC. The evaluation protocol includes both subject-dependent and subject-independent settings, and six EDL measures: Chebyshev distance, Clark distance, Canberra distance, Kullback–Leibler divergence, Cosine similarity, and Intersection similarity. On DMER subject-dependent, HeLo achieves the best performance on 5 of 6 metrics and an Average Rank = 1.16; on WESAD subject-dependent, it is best on all six metrics with Average Rank = 1.0. In the more difficult subject-independent setting, HeLo is best across all six metrics on DMER with Average Rank = 1.0, and on WESAD with Average Rank = 1.16. The paper also reports model efficiency relative to several deep baselines: HeLo uses about 11.656M FLOPs and 4.040M parameters, compared with 91.955M FLOPs and 15.672M parameters for MAET, 330.671M FLOPs and 6.780M parameters for CARAT, and 156.354M FLOPs and 0.483M parameters for EmotionDict (Zheng et al., 9 Jul 2025).

The ablation studies identify all three core modules as important. Removing CAPF, replacing OTHM with simple concatenation, or dropping LCDCA all degrades performance. Further ablations show that removing any modality harms results, that EEG is the best query modality within CAPF, that 4 attention heads perform best, and that a transformer depth of 1 outperforms deeper variants, which the paper attributes to overfitting in deeper configurations (Zheng et al., 9 Jul 2025).

4. Fed-HeLLo as federated fine-tuning with heterogeneous LoRA allocation

In "Fed-HeLLo: Efficient Federated Foundation Model Fine-Tuning with Heterogeneous LoRA Allocation" (Zhang et al., 13 Jun 2025), the name denotes a federated-learning framework for collaborative fine-tuning of foundation models under heterogeneous client resources. The central idea is to attach LoRA modules to transformer layers, freeze the backbone, and allow different clients to train different subsets of LoRA layers according to their local capacity. The low-rank update is written as

$T_n$6

with $T_n$7 and $T_n$8.

The problem formulation assumes a central server, private client datasets $T_n$9, a frozen pre-trained model $M$0, and a global trainable parameter set of LoRA modules. At round $M$1, each selected client $M$2 has a capacity $M$3, interpreted as the number of LoRA layers it can train. The server constructs an allocation map

$M$4

subject to $M$5. Active layers are trained locally; inactive layers remain frozen. Aggregation is layer-wise, over only those clients that updated a given layer: $M$6

The distinctive contribution is the family of Heterogeneous LoRA Allocation (HLA) strategies. FIM-HLA estimates dynamic layer importance through a Fisher Information Matrix proxy based on squared gradient norms: $M$7 These scores are converted into probabilistic layer-allocation distributions. GD-HLA instead encodes intrinsic layer importance through four geometric patterns across depth: Triangle, Inverted Triangle, Bottleneck, and Uniform. RGD-HLA is the randomized version of GD-HLA, using the geometric pattern to induce layer-level probabilities and then sampling client-specific active subsets without replacement. The full Fed-HeLLo design combines the geometric prior with periodic FIM-based updates.

The motivation is explicitly resource-oriented. The paper states that, for ViT-base, LoRA fine-tuning still has substantial memory cost because only around 3% of memory is model parameters while 97% is optimizer states plus activations, and that applying LoRA to every transformer layer is similar to full fine-tuning in memory footprint. Heterogeneous layer-wise allocation is therefore proposed as an alternative to either excluding weaker clients or forcing all clients to operate at the weakest client’s capacity.

Experiments cover CIFAR-100, DomainNet-121, LEDGAR, Natural Instruction, and Dolly-15K, with ViT-base, BERT-base, DataJuicer-1B, and OPT-1.3B backbones under different Non-IID conditions and two resource-heterogeneity ratios: 6:3:1 and 1:1:1. Representative results reported in the paper include the following averages under RoC 6:3:1: on CIFAR-100, Fed-HeLLo 82.52% versus FedRA 80.68%, HETLoRA 65.88%, and Straggler Learning 61.50%; on LEDGAR, Fed-HeLLo 64.54 macro-F1 versus FedRA 61.39; on Natural Instruction, Fed-HeLLo 62.48 Rouge-L versus FedRA 60.75; on Dolly-15K, Fed-HeLLo 59.71 Rouge-L versus FedRA 59.12; and on DomainNet-121, Fed-HeLLo 57.30% versus FedRA 55.34%. Under RoC 1:1:1, the method also remains competitive, including 83.84% on CIFAR-100 and 66.04 on LEDGAR. The paper further reports that, relative to full-LoRA Exclusive Learning on CIFAR-100, Fed-HeLLo reduces average backward compute from 3.25 to 2.03 TFLOPs and saves about 1.5 GB memory per step (Zhang et al., 13 Jun 2025).

Among the geometric patterns, the Bottleneck configuration is reported as the strongest in the cited GD-HLA ablation on CIFAR-100, including 77.79% average accuracy versus 70.26% for Uniform in the 100-client 6:3:1 setting. The paper interprets the combined strategy as trading off early-round stability from geometric priors against later-round adaptivity from FIM-based importance estimates (Zhang et al., 13 Jun 2025).

5. Evaluation regimes, assumptions, and limitations

The three HeLo/HELO systems differ sharply in what counts as evidence, and each paper states nontrivial constraints on interpretation.

For HELO cryptography, the experimental platform is a desktop PC, specifically an Intel Core i5-9600K system with 16 GB DDR4-2666 RAM and Windows 11 Pro, rather than a microcontroller or deployed IoT endpoint. Energy is estimated from CPU TDP and CPU time rather than measured directly. The paper therefore presents its performance as relevant to IoT-like workloads, but the hardware results are extrapolative rather than in situ. The system also performs worse than AES-based approaches for larger 20–40 MB image files, so its strongest regime is small to medium textual or structured data rather than large multimedia payloads (Ahmed et al., 5 May 2026).

For HeLo emotion distribution learning, the framework assumes synchronized multi-modal windows and access to emotion distributions as supervision. The paper explicitly notes the computational cost of OT with Sinkhorn relative to simpler fusion schemes, the reliance on synchronized physiological and behavioral data, the requirement for distributional labels such as normalized PANAS ratings, and the quadratic scaling of the $M$8 label-correlation matrix with the number of labels. The label embeddings are randomly initialized; the paper identifies LLM-based initialization as a future direction rather than a present component (Zheng et al., 9 Jul 2025).

For Fed-HeLLo, the main focus is resource heterogeneity rather than direct treatment of privacy or data-heterogeneity mitigation beyond the federated setup itself. The paper notes that federated learning remains vulnerable to model inversion and related attacks, and states that differential privacy and secure aggregation could be layered on top. It also remarks that activation storage remains heavy even with partial-layer training, and points to activation pruning, checkpointing, activation compression, model distillation, and data-heterogeneity-specific techniques as future directions (Zhang et al., 13 Jun 2025).

A second misconception is to interpret the shared name as implying shared design constraints. Only the cryptographic HELO is explicitly a lightweight security system for IoT; only the emotion-recognition HeLo is an EDL architecture centered on heterogeneous fusion and label correlation; only Fed-HeLLo addresses federated LoRA fine-tuning under client resource heterogeneity.

6. Cross-domain interpretation of the name

Although the three frameworks are unrelated in lineage, they exhibit an interpretable structural resemblance. Each rejects uniform processing in favor of structured allocation across heterogeneous components. In HELO cryptography, the structure is a hybrid composition of asymmetric and symmetric cryptography, MACs, signatures, chunking, and multithreading. In HeLo for emotion recognition, the structure lies in staged fusion of physiological and behavioral modalities and in explicit label-correlation modeling. In Fed-HeLLo, the structure is the assignment of different LoRA-layer subsets to clients with different capacities (Ahmed et al., 5 May 2026, Zheng et al., 9 Jul 2025, Zhang et al., 13 Jun 2025).

This suggests a common design intuition behind the separate names: performance and robustness are improved not by treating all channels, labels, or layers identically, but by modeling their asymmetries explicitly. The cryptographic paper operationalizes this through hybrid primitives and chunk-aware transmission; the affective-computing paper through OT-based cross-modal alignment and correlation-driven attention; and the federated-learning paper through importance-aware, capacity-constrained layer selection. A plausible implication is that the recurrence of the name “HeLo” reflects the broader appeal of methods that combine heterogeneity-awareness with efficiency constraints, even when the underlying technical objects—IoT ciphertexts, emotion distributions, and LoRA layers—are entirely different (Ahmed et al., 5 May 2026, Zheng et al., 9 Jul 2025, Zhang et al., 13 Jun 2025).