BIDO: Multifaceted Applications in Diverse Domains

• BIDO is an acronym applied to distinct constructs, including bilateral dependency optimization for MI defense, biomedical ontologies, device-free biometric authentication, bidifferential algebra, and image-based malware detection.
• Its implementations range from enhancing privacy in neural networks and structuring bacteria infectious disease data to securing online identity and advancing algebraic geometry.
• Technical details include dual-dependency regularization, OWL-based ontology design, deterministic key derivation via biometric analysis, bidifferential frameworks, and robust cross-modal feature fusion for malware detection.

In the supplied research literature, BIDO is not a single canonical technical term but a reused acronym or shorthand applied to several unrelated constructs. Its attested uses include Bilateral Dependency Optimization for defending neural classifiers against model-inversion attacks, the Bacteria Infectious Disease Ontology within the IDO ecosystem, Biometric Identity Online as a device-free WebAuthn-compatible authentication framework, a shorthand for bidifferential algebra and $B$-geometry, and a hybrid image-based malware detector designed to address obfuscation and concept drift (Peng et al., 2022, Babcock et al., 2024, Mithra et al., 16 May 2026, Sanchez et al., 2021, Li et al., 4 Sep 2025).

1. Disambiguation and research usage

The acronym is reused across machine learning, ontology engineering, authentication, algebraic geometry, and cybersecurity. In each case, the surrounding formalism is field-specific, and the meanings are not interchangeable.

Usage of BIDO	Domain	Source
Bilateral Dependency Optimization (BiDO)	Privacy-preserving deep learning	(Peng et al., 2022)
Bacteria Infectious Disease Ontology	Biomedical ontology engineering	(Babcock et al., 2024)
Biometric Identity Online	Biometric authentication and WebAuthn	(Mithra et al., 16 May 2026)
Bidifferential shorthand for commutative biderivation theory	Commutative algebra and algebraic geometry	(Sanchez et al., 2021)
Hybrid image-based malware detector	Android malware detection	(Li et al., 4 Sep 2025)

A recurrent source of confusion is that some papers use BIDO as a full acronym, whereas the bidifferential paper uses it only as contextual shorthand for a theory of commutative rings equipped with a biderivation. Another point of ambiguity is orthographic: the privacy-defense paper writes BiDO, but it denotes the same acronymic expansion, Bilateral Dependency Optimization.

2. Bilateral Dependency Optimization in model-inversion defense

In privacy-preserving deep learning, BiDO is a training strategy for defending classifiers against model-inversion (MI) attacks. The central idea is to regularize latent representations in two directions simultaneously: reduce the dependency between inputs $X$ and latent layers $Z_j$, while increasing the dependency between $Z_j$ and labels $Y$. This is intended to suppress reconstructible private information without directly undermining the supervised objective. The general objective is given as

$$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$

where $L(\theta)$ is the standard supervised loss, $d(\cdot,\cdot)$ is a dependence measure, and $\lambda_x,\lambda_y>0$ are balancing coefficients (Peng et al., 2022).

The paper contrasts this with prior unilateral dependency optimization, especially MID, which minimizes dependency between inputs and outputs directly. That earlier strategy is described as being in explicit tension with supervised learning, because classification training aims to maximize dependence between inputs and outputs. BiDO therefore shifts the regularization target from the final input-output map to the internal representation.

Two concrete realizations are proposed. BiDO-COCO uses constrained covariance (COCO), with empirical estimator

$$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$

where $X$0, $X$1, and $X$2. BiDO-HSIC uses the Hilbert-Schmidt Independence Criterion (HSIC) with empirical estimator

$X$3

The paper uses Gaussian kernels for $X$4 and $X$5, and a linear kernel for $X$6. Mini-batch computation yields complexity $X$7 for BiDO-COCO and $X$8 for BiDO-HSIC.

Evaluation is conducted in the white-box MI setting on CelebA, MNIST, and CIFAR-10, using VGG16, ResNet-34, and LeNet as target classifiers, and GMI, KED-MI, and VMI as attacks. Reported metrics include attack accuracy, Top-5 attack accuracy, FID, and defended-model classification accuracy. On CelebA against GMI, a representative BiDO-HSIC setting reports Attack Acc 6.47, Top-5 Attack Acc 16.07, and Classification accuracy 80.35, compared with MID’s 15.73, 35.27, and 78.70. On MNIST against KED-MI, BiDO-HSIC reduces attack accuracy from 42.52 to 4.36 while changing classification accuracy from 99.94% to 99.61%. The paper’s ablations indicate that both terms in the bilateral objective are necessary: removing the $X$9 penalty weakens protection, while removing the $Z_j$0 reward harms classification.

3. BIDO as the Bacteria Infectious Disease Ontology

In biomedical ontology engineering, BIDO denotes the Bacteria Infectious Disease Ontology, the bacterial “spoke” in a fourfold pathogen reference ontology suite built around the Infectious Disease Ontology (IDO) hub. The suite comprises VIDO, BIDO, MIDO, and PIDO, and follows a hub-and-spoke methodology in which IDO supplies the general infectious-disease structure and the pathogen-specific reference ontologies provide intermediate reusable layers for downstream application ontologies (Babcock et al., 2024).

BIDO is explicitly described as a reference ontology for bacterial pathogens and bacterial pathogenesis. Its scope includes bacteria and bacterial types, bacterial infections, infectious disorders, infectious diseases, bacterial pathogenesis, virulence factors, adhesion factors, toxins, bacterial populations, colonies, and host-interaction processes. The ontology is said to represent the detailed characterization of bacteria, including taxonomy, genetic composition, and reproduction mechanisms, while also covering clinical and biological host-bacteria interactions. At the same time, the paper emphasizes that BIDO is not exhaustive of bacterial biology or disease.

The ontology is engineered in OWL 2, developed in Protégé, tested with HermiT and Pellet, aligned with OBO Foundry practices, and reuses imported relations from the Relations Ontology (RO). It also reuses content from IDO, OGMS, GO, ChEBI, PRO, OMP, and CMO. Term inclusion is said to be guided by authoritative literature, expert consultation, consensus-building, and practical research needs. The paper highlights the definitional pattern “A is a B that C’s” as a governing principle for clean hierarchical definitions.

Architecturally, BIDO introduces bacterial type classes under a central bacteria class, including bacilli bacteria, cocci bacteria, spirilla bacteria, and spirochetes bacteria. Its major classes include bacterial infection, bacterial infectious disorder, bacterial infectious disease, bacterial infectious disease course, bacterial pathogenesis, bacterial adhesion disposition, bacterial adhesion factor, bacteria population, bacteria colony, appearance of bacterial infectious disorder, process of establishing a bacterial infection, and bacterial pathogenesis involving infection.

A key ontological distinction inherited from IDO is the separation between infection as the host-pathogen interaction or event and infectious disorder as the disorder arising from that process. BIDO further distinguishes bacterial infectious disorder from bacterial toxin disorder, because some bacterial diseases are caused by toxins without infection. The paper uses botulism as the illustrative case: C. botulinum may cause disease through toxin exposure even when the bacteria do not invade the host. Accordingly, BIDO does not require every bacterial pathogenesis instance to include host invasion or infection establishment.

The paper also presents explicit logical patterns. Bacterial infectious disease is defined as an infectious disease with material basis in some bacterial infectious disorder. Bacterial pathogenesis is characterized as a pathogenesis process realization of a pathogenic disposition inhering in some bacteria or bacteria population, with at least the proper process parts tixin biosynthetic process and appearance of disorder. A bacterial adhesion disposition is an adhesion disposition borne by a macromolecule part of some bacteria that enables adherence of the bacterium to a host. The ontology’s pilus axiom models a pilus as a cell projection, part of some bacteria, with antigen role, bacterial adhesion disposition, and virulence factor disposition, so that automated reasoning can infer it as both an adhesion factor and a virulence factor.

4. BIDO as Biometric Identity Online

In authentication research, BIDO stands for Biometric Identity Online, a device-free biometric authentication framework intended to authenticate a person from any commodity sensor terminal without a hardware token, smart card, pre-provisioned private key, or persistent biometric template. Its core mechanism is deterministic key derivation from a live biometric measurement plus a memorized secret, yielding a transient ECDSA P-256 key pair used for WebAuthn signing and then immediately zeroized (Mithra et al., 16 May 2026).

The reference implementation is facial. It uses OpenCV for capture, the Dlib 68-point face landmark predictor, a frontal face detector, face alignment, a frontality gate, landmark-distance vectorization, quantization, salt appending, and SHA-256 hashing. A frame is accepted only if exactly one face is present. The eye centers are computed from landmarks 36–41 and 42–47, and the face is aligned into a 200×200 crop with target eye positions at $Z_j$1 and $Z_j$2. The inter-eye midpoint $Z_j$3, distance $Z_j$4, and tilt angle $Z_j$5 determine the affine transform. After alignment, a frontality check rejects non-frontal frames when the horizontal spans of the left and right eyes relative to $Z_j$6 differ.

Feature vectorization uses distances from $Z_j$7 to a selected landmark subset $Z_j$8,

$Z_j$9

with the entropy analysis using $Z_j$0 selected landmarks. Each distance is quantized by floor division with divisor $Z_j$1,

$Z_j$2

A user-provided secret $Z_j$3 is appended, and the byte array is hashed as

$Z_j$4

Enrollment uses exactly 200 valid frames. Each frame yields a hash, and the system selects the mode of the 200 digests as the Verification Seed (Vseed): $Z_j$5 The paper describes this majority vote as a stabilization mechanism against quantization boundary crossings, inter-frame jitter, and minor pose or lighting variation. Vseed is then used to derive an ECDSA P-256 key pair transiently. A CredID is constructed by signing a fixed non-secret string Vconst with the private key and prepending a FIXED_PREFIX. The relying party stores only CredID and PubKey, and the credential is a non-discoverable / non-resident WebAuthn credential.

Authentication is lighter. It does not recompute the 200-frame majority vote. Instead, the framework reconstructs candidate hashes one frame at a time, derives a candidate key pair, and checks whether the candidate public key can verify the stored signed Vconst. If verification fails, the candidate state is zeroized and the next frame is tried; if it succeeds, the recovered private key signs the relying party’s challenge. The framework is claimed to achieve AAL2 under NIST SP 800-63B, with the biometric as one factor and the memorized secret as the second factor.

The evaluation uses VGGFace2, LFW, and MegaFace Challenge 1, with a face encoder based on ResNet-50 and a 512-d ArcFace embedding trained on VGGFace2. Reported performance includes 99.51% verification accuracy and EER 0.49% on LFW, 92.14% Rank-1 accuracy and TAR at FAR $Z_j$6 of 94.37% on MegaFace Challenge 1 with $Z_j$7 distractors, and a 128-bin + 200-vote proposed method achieving 99.1% match, Crypto-FAR 0.03%, and Crypto-FRR 0.90% on the VGGFace2 binding-consistency test. On an ARM Cortex-A53 @ 1.4 GHz with hardware-accelerated SHA-256, average end-to-end latency is reported as 191 ms with 14 ms standard deviation.

The paper states that BIDO does not persist private keys, biometric templates, facial images, or other PII. It also notes several limitations: a compromised terminal could capture transient key material before zeroization, the entropy analysis assumes independence among landmarks, highly similar faces such as monozygotic twins increase collision risk, self-signed attestation does not satisfy FIDO2 Metadata Service-verified attestation requirements, and liveness is not provided by Dlib itself. For presentation attack detection, the system integrates a separate ISO/IEC 30107-3 PAD module with reported BPCER 2.1% and APCER 3.8%.

5. BIDO as bidifferential algebra and $Z_j$8-geometry

In commutative algebra and algebraic geometry, BIDO is used as shorthand for the bidifferential framework developed in “Commutative bidifferential algebra.” The paper studies commutative unitary rings equipped with a biderivation, meaning a binary operation $Z_j$9 that is a derivation in each argument. Unlike Poisson algebraic geometry, the setup does not assume skew-symmetry or the Jacobi identity unless explicitly stated (Sanchez et al., 2021).

A bidifferential ring $Y$0 therefore satisfies the Leibniz rule in both variables: $Y$1 For each $Y$2, the derivations $Y$3 and $Y$4 are called the associated hamiltonians. The paper gives the example

$Y$5

with derivations $Y$6 and coefficients $Y$7. Every Poisson bracket is therefore a biderivation, but not every biderivation is Poisson.

The algebraic theory includes bidifferential ideals, defined by $Y$8 and $Y$9, and bidifferential morphisms preserving the bracket. The quotient by a bidifferential ideal carries a unique induced biderivation. To handle extensions, the paper introduces a three-sorted formalism with rings $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$0, an additive group $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$1 that is both an $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$2-module and an $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$3-module, and a biderivation $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$4. This permits systematic treatment of localization, algebraic extension, transcendental extension, and tensor products.

A major part of the paper establishes extension results. Biderivations extend uniquely through localization and are lifting. For algebraic extensions, if $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$5 is separably algebraic over $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$6, then one localizes at

$$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$7

where $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$8 is the minimal polynomial of $$L(\theta) + \lambda_x \sum_{j=1}^{M} d(X,Z_j) - \lambda_y \sum_{j=1}^{M} d(Z_j,Y),$$9. As a special case, every biderivation on a field extends uniquely to its separable closure. If $L(\theta)$0 is transcendental, biderivations extend uniquely to $L(\theta)$1 with prescribed values on the new variable; in particular, every biderivation extends to the polynomial ring $L(\theta)$2.

The geometric theory is built from $L(\theta)$3-varieties over a bidifferential field $L(\theta)$4 of characteristic zero. A $L(\theta)$5-variety is an affine algebraic variety $L(\theta)$6 together with a biderivation on $L(\theta)$7 extending that on $L(\theta)$8. A $L(\theta)$9-subvariety is defined by a bidifferential vanishing ideal, a $d(\cdot,\cdot)$0-point is a $d(\cdot,\cdot)$1-rational point whose singleton is a $d(\cdot,\cdot)$2-subvariety, and a $d(\cdot,\cdot)$3-morphism is a regular map with bidifferential pullback. Compatible base extension is developed so that $d(\cdot,\cdot)$4-subvarieties remain $d(\cdot,\cdot)$5-subvarieties after base change; for the algebraic closure $d(\cdot,\cdot)$6, every $d(\cdot,\cdot)$7-variety is shown to have a unique compatible $d(\cdot,\cdot)$8-structure on $d(\cdot,\cdot)$9.

The paper’s geometric highlight is the existence of generic $\lambda_x,\lambda_y>0$0-fibres. For a dominant $\lambda_x,\lambda_y>0$1-morphism $\lambda_x,\lambda_y>0$2 of irreducible $\lambda_x,\lambda_y>0$3-varieties, it proves the existence of a nonempty open $\lambda_x,\lambda_y>0$4 with suitable $\lambda_x,\lambda_y>0$5-structure on $\lambda_x,\lambda_y>0$6, and then deduces that every dominant $\lambda_x,\lambda_y>0$7-morphism has a generic $\lambda_x,\lambda_y>0$8-fibre over $\lambda_x,\lambda_y>0$9. This construction underpins a universal base extension obtained by iterating the generic-fibre construction.

The final section formulates the bidifferential Dixmier–Moeglin equivalence problem. For $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$0, the paper defines $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$1-locally closed, $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$2-primitive, and $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$3-rational prime bidifferential ideals and proves the implications

$$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$4

and, when $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$5 lies in the constants of $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$6,

$$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$7

The central question is then: for which affine bidifferential $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$8-algebras does every $$\widehat{COCO}(X,Y) = \frac{1}{N}\sqrt{\left\|\widetilde{K}\widetilde{L}\right\|_2},$$9-rational prime bidifferential ideal become $X$00-locally closed? This is the paper’s BDME problem.

6. BIDO as a robust image-based malware detector

In Android malware detection, BIDO is a hybrid image-based malware detector proposed to address obfuscation and concept drift as two manifestations of out-of-distribution (OOD) shift. The paper’s central claim is that prior work typically treats them as separate problems, even though both arise when the test distribution differs from the training distribution. BIDO is designed to improve both discriminability and robustness under these shifts (Li et al., 4 Sep 2025).

The pipeline has five modules: DEX-image generation, XML-image generation, local feature selection, cross-modal dependency modeling, and learnable measurement. For the DEX image, the method converts classes.dex into an RGB image but retains only the index section, discarding the header and data section on the grounds that the index is more informative for malware-relevant patterns. The XML branch converts AndroidManifest.xml directly into an RGB image from raw hex without manual feature engineering. These produce

$X$01

The local feature selection module is applied to the DEX image. A fine-tuned Inception V3 yields a Mixed-6e feature map $X$02, and a $X$03 convolution produces a mask

$X$04

whose entries estimate the likelihood that a location contains discriminative malware information. Local feature maps are then extracted via

$X$05

A self-attention mechanism with a learnable classification token and positional embedding forms the DEX representation $X$06.

The cross-modal dependency modeling stage fuses DEX and XML representations by an outer product space (OPS). The XML image is encoded by a fine-tuned ResNet and MLP to give $X$07. Their dependency matrix is

$X$08

so each element captures explicit pairwise co-occurrence, $X$09. Because the full OPS can be noisy and high-dimensional, the method uses a factorization technique from Gao et al. to compress the representation into $X$10.

The final component is a learnable Mahalanobis metric

$X$11

optimized with a contrastive loss that pulls same-label samples together and pushes different-label samples apart: $X$12 The full training loss combines three cross-entropy terms and the contrastive term,

$X$13

with $X$14, $X$15, $X$16, and $X$17.

Evaluation uses two datasets. Data-Ideal is built from APKs from Google Play, Androzoo, and CICMalDroid2020, with 12,375 malicious and 12,455 benign samples after filtering and balancing. Data-Obfu is produced by applying six obfuscation techniques—ClassRename/MethodRename, ResStringEncryption, Control Flow Obfuscation, NewAlignment, NewSignature, and Junk Code Insertion—yielding 12,088 malicious and 11,044 benign samples. Baselines are DexRay, Dex-CNN, MADRF-CNN, XMal, and Malscan. Training uses an NVIDIA RTX 3090, an 80/10/10 split, $X$18 local feature maps, 64 epochs, batch size 8, and SGD with momentum 0.9.

On unobfuscated detection (RQ1), BIDO reports Accuracy 94.15%, Precision 94.17%, Recall 94.23%, and F1 94.20% on Data-Ideal. Under obfuscation (RQ2), the method remains the best among the reported models, with Accuracy 93.98%, Precision 95.70%, Recall 92.54%, and F1 94.09% in the laboratory Data-Obfu setting, and Accuracy 92.30%, Precision 93.36%, Recall 91.56%, F1 92.45% in the 45% Data-Ideal + 45% Data-Obfu training scenario. Under concept drift (RQ3), training on 2016 samples and testing on 2017 samples yields Accuracy 84.40%, Precision 81.16%, Recall 89.60%, and F1 85.17%, again the strongest reported performance.

The ablations attribute the gain to the combination of modalities and fusion method. Model-xml alone reaches F1 89.40, Model-dex reaches F1 92.70, and full BIDO reaches F1 94.20, indicating complementarity between XML and DEX. Among fusion strategies, OPS outperforms CrossAttention, Summation, and Concatenation, and the number of local feature maps peaks at $X$19. Within this usage, BIDO is therefore a specifically OOD-oriented detector rather than a generic image classifier: its architectural choices are tied to the paper’s claim that both obfuscation and concept drift should be addressed through stable cross-modal dependencies and compact class geometry.