Skip to the content.

Annotation Schema

← Back to Home

← Data Preprocessing

This page describes the metadata and annotation schema used to classify the primary studies in the survey Fuzzing AI Systems: Foundations, Techniques, and Open Challenges.

After the final study-selection process, each included study was represented using both bibliographic metadata and technical annotation fields. The schema was designed to support the research questions and enable consistent comparison across different AI-system targets, fuzzing techniques, oracle mechanisms, and failure types.

Bibliographic Metadata

The following bibliographic fields were retained for each selected study.

Field Description
Paper ID Unique identifier assigned to the study
Title Full title of the publication
Authors Authors listed in the publication
Publication year Year associated with the selected publication version
Venue Conference, journal, workshop, or preprint venue
Publisher Publisher or publication platform
DOI Digital Object Identifier, when available
URL Publisher, repository, or publication link
Abstract Abstract text used during screening and verification

Technical Annotation Fields

The technical schema contained 17 fields.

No. Annotation field Purpose
1 Testing Target Identifies the direct artifact exercised by the fuzzing approach
2 System Level Identifies the AI-system layer at which behavior or failure is observed
3 AI Paradigm Records the AI paradigm addressed by the study
4 Target Framework/Platform Records the framework, library, compiler, backend, simulator, or platform under test
5 Technique Name Records the name of the proposed technique, tool, or fuzzing approach
6 Technique Family Classifies the high-level fuzzing or testing mechanism
7 Input Generation Strategy Describes how initial test inputs are generated or selected
8 Mutation Strategy Describes how existing inputs, programs, models, prompts, or scenarios are transformed
9 Oracle Type Identifies the high-level mechanism used to determine whether a failure occurred
10 Oracle Construction Describes how the oracle is implemented or derived
11 Failure Type Records the type of failure exposed by the approach
12 Reported Limitations Captures limitations explicitly reported by the study
13 Threats to Validity Records threats to validity discussed by the authors
14 Future Work Captures future research directions proposed by the study
15 Observed Gap Records gaps identified through evidence-based review
16 Main Contribution Summarizes the primary contribution of the study
17 Notes Stores supporting evidence, clarification, or annotation comments

Main Classification Dimensions

Testing Target

The testing target identifies the concrete artifact directly exercised by the fuzzer.

The main categories include:

Some studies may span more than one target layer.

System Level

The system-level field captures the layer at which behavior or failure is observed.

This distinction is useful because the direct testing target and the level at which the failure becomes visible are not always identical.

AI Paradigm

This field records the primary AI paradigm addressed by the study, such as:

Target Framework or Platform

This field records the concrete software ecosystem or execution platform examined by the study.

Examples may include:

Technique Family

The technique-family field captures the high-level testing or fuzzing mechanism used by a study.

The main technique families include:

Technique families are not necessarily mutually exclusive. A study may combine several mechanisms within a hybrid workflow.

Input Generation Strategy

This field records how the initial test inputs are produced or selected.

Examples include:

Mutation Strategy

This field records how existing inputs or test artifacts are modified.

Examples include:

Input generation and mutation were recorded separately because studies may use the same high-level technique family while differing substantially in how tests are created and transformed.

Oracle Type

The oracle-type field captures the high-level mechanism used to determine whether the observed behavior indicates a failure.

The main oracle types include:

Oracle Construction

Oracle construction describes how the failure-detection mechanism is implemented.

Common construction mechanisms include:

Oracle type and oracle construction were recorded separately because the same oracle type may be implemented in different ways.

Failure Type

The failure-type field records the form of abnormal behavior exposed by the fuzzing approach.

The main categories include:

Failure categories were treated as multi-label when a study reported more than one type of failure.

Multi-Label Annotation

Several annotation dimensions were treated as non-mutually exclusive.

A study could receive multiple labels when it:

This design preserved the hybrid and target-dependent nature of AI-system fuzzing approaches rather than forcing each study into a single category.

Schema Development

The annotation schema was developed from the research questions and refined iteratively during pilot annotation.

Controlled labels were used for the primary classification fields to improve consistency, while evidence notes were retained for ambiguous or complex cases.

The schema supported the subsequent LLM-assisted annotation, human verification, consistency checking, and final synthesis reported in the survey.

Further details are available on the LLM-Assisted Annotation page and the Annotation Quality Control page.