strawberry/RESEARCH.md
2026-04-26 16:37:36 -04:00

286 lines
13 KiB
Markdown

# Research Brief
Working research brief for Strawberry's current approach and implementation direction.
## Linked Documents
- [PLAN.md](PLAN.md)
- [AGENTS.md](AGENTS.md)
- [README.md](README.md)
- [UX.md](UX.md)
## Scope
Strawberry's intended design combines:
- Linux runtime telemetry
- eBPF/libbpf-based process observation
- YARA-based file or process-memory matching
- graph or provenance-style relationship analysis
- a local embedded store for indicators and relationships
This is not a formal systematic literature review. It is a focused, implementation-oriented reading list assembled on **April 25, 2026** from paperbridge results to guide architecture and sequencing.
The brief was refreshed on **April 26, 2026** to sharpen two now-immediate topics:
- a real **correlation engine** for detections over Strawberry's stored process/file/indicator graph
- a real **analyst UX**, likely a native process-centric desktop app rather than only a minimal local web trace viewer
## Core Research Position
The literature supports the overall direction:
1. **Runtime process telemetry is a strong foundation**
- eBPF-based observation is now a practical basis for low-overhead Linux runtime security systems.
2. **Pure signatures are insufficient on their own**
- YARA remains useful for triage and known-pattern matching, but stealthy or fileless behavior requires behavioral context.
3. **Provenance and graph analysis materially improve detection**
- Process/file/network relationships help surface multi-step attacks that point detectors miss.
4. **Online graph systems must be concise and efficient**
- The strongest practical systems prioritize compact graph construction, bounded search, and interpretable attack paths.
5. **Provenance systems can be evaded**
- Mimicry and behavior blending are real risks, so Strawberry should treat graph analysis as one layer, not a complete defense.
6. **Operators need correlated findings, not raw graph exhaust**
- The literature on threat hunting and SOC workflows repeatedly points toward aggregation, explanation, and human-usable pivots rather than standalone low-level events.
## Implications for Strawberry
### 1) eBPF telemetry should come before advanced graph analytics
Without trustworthy runtime events, the graph layer is mostly hypothetical. The literature consistently assumes reliable event capture first, then higher-order reasoning.
### 2) YARA should be integrated as an enrichment layer
The best fit for Strawberry is:
- use eBPF or process telemetry to decide **what to inspect**
- use YARA to scan files or memory regions when warranted
- store both the match and the surrounding execution context
### 3) The embedded database should preserve relationships, not just hits
A local store is most useful when it keeps:
- indicators
- process metadata
- parent/child edges
- file access or execution links
- timestamps
This makes later provenance or graph-style analysis possible even before advanced ML is added.
### 4) The first milestone should be deterministic, local, and inspectable
The literature shows the value of online systems, but Strawberry is earlier than that. The right near-term target is a small, verifiable local pipeline:
1. capture a process event
2. resolve an executable path
3. hash and/or scan it
4. persist the result and relationship edges
5. query the stored trail back out
### 5) The next major subsystem should be correlation
Strawberry now has enough substrate to justify a correlation layer:
- process execution events
- process lineage edges
- executable-image file nodes
- YARA-backed indicator hits
- timestamps and partial runtime context
The next literature-aligned step is to convert these into **finding candidates** or **incident summaries** rather than leaving them as raw stored graph elements.
### 6) The long-term operator surface should be process-centric
The current local web GUI is good for evaluation, but the operator workflow implied by the literature is:
- search for a process, executable, or indicator
- pivot through lineage and related artifacts
- inspect a concise correlated explanation
- expand into supporting evidence only when needed
That argues for a richer analyst surface. A native Qt app is a reasonable fit if Strawberry stays local-first and desktop-oriented.
## Literature by Theme
### A. Provenance / graph-based detection
#### GID: Graph-based Intrusion Detection on Massive Process Traces for Enterprise Security Systems
- Boxiang Dong et al., 2016
- Link: https://huggingface.co/papers/1608.02639
- Why it matters:
- Strong evidence that graph representations of heterogeneous process traces improve detection over single-event views.
- Relevant to Strawberry's process + relationship model.
#### You Are What You Do: Hunting Stealthy Malware via Data Provenance Analysis
- Qi Wang et al., 2020
- DOI: https://doi.org/10.14722/ndss.2020.24167
- Why it matters:
- Strong theoretical and practical support for provenance-centric malware hunting.
- Reinforces that behavior chains can expose stealthy attacks even when obvious payload artifacts are absent.
#### Unicorn: Runtime Provenance-Based Detector for Advanced Persistent Threats
- Xueyuan Han et al., 2020
- DOI: https://doi.org/10.14722/ndss.2020.24046
- Why it matters:
- Demonstrates that long-lived, low-and-slow attacks can be modeled from runtime provenance.
- Relevant to Strawberry's interest in persistent local graph storage.
#### NODLINK: An Online System for Fine-Grained APT Attack Detection and Investigation
- Shaofei Li et al., 2024
- DOI: https://doi.org/10.14722/ndss.2024.23204
- Why it matters:
- Highlights the importance of concise, interpretable online provenance graphs.
- Useful design pressure against building an overgrown, low-signal graph store.
#### POIROT: Aligning Attack Behavior with Kernel Audit Records for Cyber Threat Hunting
- Sadegh M. Milajerdi et al., 2019
- Link: https://huggingface.co/papers/1910.00056
- Why it matters:
- Shows how graph matching over kernel-audit-derived provenance can support threat hunting.
- Suggests a future direction for query-driven hunt workflows in Strawberry.
#### Sometimes, You Aren't What You Do: Mimicry Attacks against Provenance Graph Host Intrusion Detection Systems
- Akul Goyal et al., 2023
- DOI: https://doi.org/10.14722/ndss.2023.24207
- Why it matters:
- Important cautionary paper.
- Strawberry should assume provenance graphs are valuable but attackable; YARA, hashing, and policy signals should remain complementary layers.
### B. YARA and rule-based enrichment
#### Scanning memory with Yara
- Michael Cohen, 2017
- DOI: https://doi.org/10.1016/j.diin.2017.02.005
- Why it matters:
- Directly relevant to Strawberry's proposed process-memory scanning path.
- Supports treating YARA as more than just a file scanner.
#### APIARY: An API-based automatic rule generator for YARA to enhance malware detection
- Antonio Coscia et al., 2025
- DOI: https://doi.org/10.1016/j.cose.2025.104397
- Why it matters:
- Not required for v1, but useful for future rule-authoring workflows.
- Suggests Strawberry could later support generated or curated rule pipelines instead of only hand-maintained rules.
### C. eBPF and Linux runtime security
#### bpfbox
- William Findlay, Anil Somayaji, David Barrera, 2020
- DOI: https://doi.org/10.1145/3411495.3421358
- Why it matters:
- Demonstrates practical eBPF-based confinement and security instrumentation on Linux.
- Useful proof that eBPF can anchor security controls, not just passive observation.
#### Comparative Analysis of eBPF-Based Runtime Security Monitoring Tools in Monitoring and Threat Detection on Kubernetes
- Aldien Asy Syairozi and Arizal, 2025
- DOI: https://doi.org/10.5220/0014272700004928
- Why it matters:
- Practical comparative context for modern eBPF-based security tooling.
- Even though Kubernetes is not Strawberry's immediate scope, the operational tradeoffs are relevant.
#### DySec: A Machine Learning-based Dynamic Analysis for Detecting Malicious Packages in PyPI Ecosystem
- Sk Tanzir Mehedi et al., 2025
- Link: https://huggingface.co/papers/2503.00324
- Why it matters:
- Shows eBPF telemetry being used in a practical detection pipeline for malicious install-time behavior.
- Useful evidence for a staged Strawberry pipeline: observe first, enrich second, classify or score later.
### D. Correlation, threat hunting, and analyst workflow
#### Real-time Analytics for APT Detection and Threat Hunting Using Cyber-threat Intelligence and Provenance Graphs
- V.N. Venkatakrishnan, 2025
- DOI: https://doi.org/10.1145/3716815.3729016
- Why it matters:
- Direct support for a **correlation engine** that combines provenance structure with higher-level threat knowledge.
- Reinforces that Strawberry's graph should feed hunt and triage workflows, not merely archive events.
#### LogKernel: A Threat Hunting Approach Based on Behaviour Provenance Graph and Graph Kernel Clustering
- Jiawei Li et al., 2022
- Link: https://arxiv.org/abs/2208.08820
- Why it matters:
- Supports the idea that suspicious behavior should be identified over **subgraphs or paths**, not isolated events.
- Useful pressure toward compact scoring over bounded behavior windows.
#### threaTrace: Detecting and Tracing Host-based Threats in Node Level Through Provenance Graph Learning
- Su Wang et al., 2021
- Link: https://arxiv.org/abs/2111.04333
- Why it matters:
- Suggests Strawberry should preserve enough per-entity context to support future node-level scoring.
- Also supports a process-centric analyst UX where an operator starts from one suspicious entity and expands outward.
#### TDLens: Toward an empirical evaluation of provenance graph-based approach to cyber threat detection
- Rui Mei et al., 2022
- DOI: https://doi.org/10.23919/jcc.2022.00.028
- Why it matters:
- Useful reminder to validate whether the graph layer actually improves detection quality instead of assuming it does.
- Good evidence for benchmarking a future Strawberry correlation engine against simpler baselines.
#### Security Operations Center: A Systematic Study and Open Challenges
- Manfred Vielberth et al., 2020
- DOI: https://doi.org/10.1109/access.2020.3045514
- Why it matters:
- Reinforces that analyst workflow, process, and explanation matter as much as raw technical detection.
- Supports building Strawberry around concise pivots, triage views, and explainable outputs.
#### An Interview Study on Third-Party Cyber Threat Hunting Processes in the U.S. Department of Homeland Security
- William P. Maxam and James C. Davis, 2024
- Link: https://arxiv.org/abs/2402.12252
- Why it matters:
- Indicates that threat hunting workflows are still variable and tool support remains a practical pain point.
- Supports building Strawberry's next UX around fast search, pivoting, and repeatable analyst workflows.
#### Enhancing Cyber Threat Hunting — A Visual Approach with the Forensic Visualization Toolkit
- Jihane Najar et al., 2025
- Link: https://arxiv.org/abs/2509.09185
- Why it matters:
- More tentative and newer than the other references, but directionally useful for UI thinking.
- Supports graph-aware visual exploration once Strawberry's underlying search and correlation model is stable.
## Research-Grounded Design Guidance
Based on the papers above, Strawberry should prefer:
- **event capture before advanced scoring**
- **local persistence before distributed orchestration**
- **graph context plus signatures**, not either one alone
- **compact, queryable provenance trails** over vague “graph database” ambitions
- **layered detection**, because provenance-only systems can be evaded
- **correlated findings over raw hits**, because operators need explanation and prioritization
- **process-first investigation UX**, because that is the most defensible current pivot in Strawberry's data model
## Open Questions
1. Should Strawberry target:
- a lightweight host detector, or
- a threat-hunting substrate with replayable provenance?
2. Is the immediate “make” goal:
- GNU Make wrapper support, or
- Make.com automation/integration?
3. How much provenance is enough for v1?
- process spawn edges only
- process + file edges
- process + file + network edges
4. Should the first correlation engine be:
- simple rule/heuristic based
- subgraph/pattern based
- anomaly scored over bounded provenance windows
5. Should the first serious analyst UI remain web-local, or move to a native Qt desktop shell?
6. When should Strawberry add thread-level pivots?
- only after true thread/TID telemetry exists
- or via a process-first UI with deferred thread support
## Recommended v1 Research Interpretation
The best literature-aligned v1 is:
- Linux-only
- eBPF-backed process execution telemetry
- YARA-backed file and optionally process-memory scanning
- local embedded persistence
- process lineage and file relationship edges
- simple rule- or heuristic-driven **correlation and triage**
- a process-centric operator UX, with thread-level pivots deferred until the telemetry model actually supports them
That scope is ambitious enough to be meaningful and small enough to make real.