Upload cleaned dataset (Tigress runtime stripped, 980K samples, ~955 MB)

README.md

@@ -0,0 +1,180 @@
+---
+license: mit
+task_categories:
+  - translation
+language:
+  - en
+tags:
+  - assembly
+  - deobfuscation
+  - llvm-ir
+  - aarch64
+  - binary-analysis
+  - reverse-engineering
+  - code
+  - obfuscation
+  - tigress
+  - exebench
+pretty_name: Obfuscated ExeBench (AArch64)
+size_categories:
+  - 100K<n<1M
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: "data/train/*.parquet"
+---
+
+# Obfuscated ExeBench (AArch64)
+
+A large-scale dataset of **obfuscated AArch64 assembly** functions paired with their **clean LLVM IR**, original **C source code**, and **clean assembly**.  Designed for training neural models that can **deobfuscate** and **lift** obfuscated binary code.
+
+## Dataset Summary
+
+| Property | Value |
+|---|---|
+| **Total samples** | ~980,000 |
+| **Source** | [jordiae/exebench](https://huggingface.co/datasets/jordiae/exebench) (`clang` revision, `train_synth_compilable` split) |
+| **Architecture** | AArch64 (ARM64) |
+| **Obfuscator** | [Tigress 4.0.11](https://tigress.wtf/) |
+| **Compiler** | `aarch64-linux-gnu-gcc 15.2.0` (`-S -O0 -std=c11 -w`) |
+| **Techniques** | Control-Flow Flattening, Arithmetic Encoding, Combined |
+| **Format** | Parquet with Snappy compression |
+
+## Columns
+
+| Column | Type | Description |
+|---|---|---|
+| `fname` | `string` | Function name |
+| `func_def` | `string` | Original C source code of the function |
+| `technique` | `string` | Obfuscation technique applied (see below) |
+| `clean_asm` | `string` | Clean AArch64 assembly from ExeBench (`angha_gcc_arm_O0`) |
+| `obfuscated_asm` | `string` | Obfuscated AArch64 assembly (after Tigress → GCC) |
+| `clean_ir` | `string` | Clean LLVM IR from ExeBench (`angha_clang_ir_O0`) |
+| `obfuscated_c` | `string` | Tigress-obfuscated C source (target function only, runtime stripped) |
+| `tigress_seed` | `int32` | Random seed used for Tigress (for reproducibility) |
+| `exebench_split` | `string` | Source ExeBench split name |
+
+## Obfuscation Techniques
+
+Each function is independently obfuscated with one of three Tigress transformation pipelines:
+
+| Technique | Tigress Flags | Description |
+|---|---|---|
+| `Flatten` | `--Transform=Flatten` | **Control-Flow Flattening (CFF)** — replaces structured control flow with a dispatcher loop and switch statement, obscuring the original program structure. |
+| `EncodeArithmetic` | `--Transform=EncodeArithmetic` | **Arithmetic Encoding** — replaces simple arithmetic and boolean operations with equivalent but complex expressions using Mixed Boolean-Arithmetic (MBA) identities. |
+| `Flatten+EncodeArithmetic` | `--Transform=Flatten --Transform=EncodeArithmetic` | **Combined** — applies both CFF and arithmetic encoding for maximum obfuscation. |
+
+The techniques are applied in roughly equal proportions (~⅓ each).
+
+## Intended Uses
+
+### Primary: Neural Deobfuscation / Lifting
+Train sequence-to-sequence models that translate **obfuscated assembly → clean LLVM IR**:
+- Input: `obfuscated_asm`
+- Target: `clean_ir`
+
+### Secondary: Obfuscation Detection / Classification
+Build classifiers that identify which obfuscation technique was applied:
+- Input: `obfuscated_asm` or `obfuscated_c`
+- Target: `technique`
+
+### Other Uses
+- **Source-level deobfuscation**: `obfuscated_c` → `func_def`
+- **Assembly comparison studies**: `clean_asm` vs `obfuscated_asm`
+- **Benchmarking binary analysis tools** against known obfuscation
+- **Understanding Tigress transformations** at the assembly level
+
+## Quick Start
+
+```python
+from datasets import load_dataset
+
+# Stream (recommended for large datasets)
+ds = load_dataset("leachl/obfuscated-exebench", streaming=True, split="train")
+
+for sample in ds:
+    print(sample["fname"])
+    print(sample["technique"])
+    print(sample["obfuscated_asm"][:200])
+    print(sample["clean_ir"][:200])
+    break
+```
+
+```python
+# Load fully into memory (requires ~1 GB RAM)
+ds = load_dataset("leachl/obfuscated-exebench", split="train")
+print(f"Total samples: {len(ds)}")
+```
+
+## Generation Pipeline
+
+```
+ExeBench (C source + synth_deps)
+         │
+         ▼
+┌─────────────────┐
+│   Tigress 4.0.11│  ← Obfuscation at C source level
+│   --Env=aarch64  │
+│   --Transform=…  │
+└────────┬────────┘
+         │ obfuscated C
+         ▼
+┌─────────────────────────┐
+│ aarch64-linux-gnu-gcc   │  ← Cross-compilation to AArch64 asm
+│ -S -O0 -std=c11 -w     │
+└────────┬────────────────┘
+         │ obfuscated assembly
+         ▼
+┌─────────────────────────────────┐
+│  Paired with ExeBench originals │
+│  clean_asm, clean_ir, func_def  │
+└─────────────────────────────────┘
+```
+
+Each ExeBench function is processed through up to 3 obfuscation pipelines independently. Functions that fail Tigress transformation or GCC cross-compilation are silently skipped (~50% Tigress failure rate is expected).
+
+## Tigress Runtime
+
+The Tigress obfuscator prepends a large runtime (~7,400 lines / ~480 KB) to every output file. This runtime is **stripped** from the `obfuscated_c` column to save space — only the obfuscated target function body is stored.
+
+A representative copy of the full Tigress runtime is available in [`tigress_runtime.c`](tigress_runtime.c). If you need to recompile the obfuscated C, prepend this runtime to the `obfuscated_c` field.
+
+## Dataset Statistics
+
+- **Source split**: `train_synth_compilable` from ExeBench
+- **Technique distribution**: ~⅓ Flatten, ~⅓ EncodeArithmetic, ~⅓ Flatten+EncodeArithmetic
+- **Obfuscated assembly is always different** from clean assembly (verified)
+- **94%** of obfuscated functions produce longer assembly than the original
+- **Clean IR** is valid LLVM IR (100% verified on sampled subset)
+
+## Limitations
+
+- Only AArch64 architecture (no x86, RISC-V, etc.)
+- Only `train_synth_compilable` split from ExeBench (not `train_real_compilable`)
+- Only 3 Tigress transformations (Tigress supports many more: Virtualize, JIT, etc.)
+- Compiled at `-O0` only (no optimisation levels)
+- The `obfuscated_c` field contains only the target function — helper functions and the Tigress runtime are stripped
+- ~50% of ExeBench functions fail Tigress processing (complex dependencies, unsupported constructs)
+
+## Citation
+
+If you use this dataset, please cite:
+
+```bibtex
+@misc{obfuscated-exebench-2026,
+  title={Obfuscated ExeBench: A Large-Scale Dataset for Neural Deobfuscation},
+  author={Leach, L.},
+  year={2026},
+  howpublished={\url{https://huggingface.co/datasets/leachl/obfuscated-exebench}},
+}
+```
+
+### Related Work
+
+- **ExeBench**: Armengol-Estapé, J., et al. (2024). [ExeBench: an ML-scale dataset of executable C functions](https://huggingface.co/datasets/jordiae/exebench).
+- **Tigress**: Collberg, C. (2024). [The Tigress C Diversifier/Obfuscator](https://tigress.wtf/).
+
+## License
+
+This dataset is released under the [MIT License](LICENSE). The underlying ExeBench data is subject to its own license terms.

data/train/shard-00000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b4b265becc2d9ca04e7500676d993809e7d2b6c31d7ff765e674f8a9f20e61a3
+size 10237249

data/train/shard-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:660094efc10b00a20f460458b34cb824f163b3563e8483449b0f5eeeb83f431c
+size 10271530

data/train/shard-00002.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20bf472ab507086027f597432ce28e59fe77866506baa8725c23ff8d3acdcf0b
+size 10140565

data/train/shard-00003.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bd57d3c37f71fab1d27f93dcb29cf41e49ed56caeb881532f5fb734d7a1b261
+size 10134035

data/train/shard-00004.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92ae004256327c615fdfdd209eafd79f6ba5de4776bf8a0c27739b34187a4624
+size 10186701

data/train/shard-00005.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c4dee50a0500dea4e076e26d896ca13abce78bcb9b81de1d3fbfdb01acfca90
+size 10088167

data/train/shard-00006.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bf2bee558adddb220071f6aa590f6f00ca18e73b39dd772bfe1f11c94e9ddea
+size 10032606

data/train/shard-00007.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b42f200349dea45d8322ab8200efd8b7be64da892e116d045bd686a34e78f6f7
+size 10444501

data/train/shard-00008.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73d0d9adf24211f5bb02e44aa2f48cc5b1bf402ce9eb2b677ff1021cda759630
+size 10436411

data/train/shard-00009.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:16fd197aa71249f671fad5d8ace487e1e656216df2b21f711a8d513fbd72a4ca
+size 10358444

data/train/shard-00010.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9646a8e5a4fe0b485226c7e31d19169c605bb57a370fe152c96ebde13247be9
+size 10340751

data/train/shard-00011.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f91e3957ead64339e80756c15b5821c8678812cd2da131cfa615119db5f951d
+size 10207420

data/train/shard-00012.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf5921df35b11963e4229b4b5c27a6d5ff4686e0910870026ebb3d73ec463bd0
+size 10344375

data/train/shard-00013.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3566ac26a7573cecce285e0e28b046cd0489f92733fb33c082902f3e9d80019f
+size 10561211

data/train/shard-00014.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f8da9c9c388c4073d3cfa71d9494a126f470d5c04288e5230dc058996819e32e
+size 10072277

data/train/shard-00015.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:736c905011f191f4f67323499fec771b47cddab172786fcf51c4bcdf1d3b3fcf
+size 10341071

data/train/shard-00016.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bbb0fb82f9fc0e02a8b9961cec29322cb617a133a74c2b1c76b2de864517ec7
+size 10030805

data/train/shard-00017.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc79c43625989e35e0e63b20c0756c672498b9f17b7d707984bbeccd38c33857
+size 10145497

data/train/shard-00018.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dbba8d18cfa87e3776a9f1b2d27966c20ce0d02ee87c9cbf58277a0d6f6b1bbb
+size 9944070

data/train/shard-00019.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ad2b0f88a32bfa9dee671096db1626aa639978504e69948e17efae36d323743
+size 10285007

data/train/shard-00020.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:538a2f28f23a4c0b0f838e3c881f807a7e7daaeedff6bee35406c5d00e774091
+size 10381619

data/train/shard-00021.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9272f7c5b1bbf3ca85b0258b98d9f6cbe7a0ed501eeccaf072047f69581272eb
+size 10183313

data/train/shard-00022.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4fb3ddc7bcc6b4f1790d9634710d0737934870bad6f708d148d63d45de6a4d2f
+size 10036737

data/train/shard-00023.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fc4a46cb817f7493f2e5d7b37dd13fb34bbfd7a3f3685393e3d6f1c542f93d6
+size 10252865

data/train/shard-00024.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:de95480ed9f08626fbcac5cd085e26bda960953d846574dd544a2d5d510543ed
+size 10563812

data/train/shard-00025.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6fd1e0072bfbefe0cbf8c55234258e4e9668433b15dcc9410f137926023668b
+size 10293638

data/train/shard-00026.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:772179425c44a47cae8eeef0b7ff29a06beaa3520375ba9c60fd4b437bde343e
+size 10255224

data/train/shard-00027.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d352091aa9b65e4797751a377b56ad79303b775547dfb2660b47d76ab6c42cfc
+size 10326075

data/train/shard-00028.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04a29de8eb87da684b5c8a9836f4e65db7cef2aeca6c82168c5d992861897e93
+size 10252586

data/train/shard-00029.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80be313b7c3765b07446f8077d0379519ace2a761bbed6df6536d339994bdc34
+size 10114223

data/train/shard-00030.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33f650a9ef0b66a1c81960c898f89fb5765ca66d1da6f0c090d44423e756f8af
+size 10271457

data/train/shard-00031.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:121ec877e6c9c9c421d46a1df85b91cb5f03638b2ee920c390dc6598f27982bf
+size 10215428

data/train/shard-00032.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7d56022467472b6b99821ff4314666b348c697fb3138a8c382cdd9d1fec7991
+size 10153775

data/train/shard-00033.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9a6726bc5e2205030efddfd2b726ca63574706ab5cf8e9ad1778df8665f546fc
+size 10162029

data/train/shard-00034.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17f062ece8efadcce899f696d620d6a8a49ae9234d9db3ce5435d195e616d536
+size 10100748

data/train/shard-00035.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1f381cf07ab91da1a5c6c6e3503a59fd85b98a389bf659b55bc742741421344c
+size 10204412

data/train/shard-00036.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ae805fe9abfe63b3feda25c93a8172d3b9bac038ca95d2a35e00f7aa2755eb4
+size 10087589

data/train/shard-00037.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e026b429b3b84b311f509392a585b5d4ea4664121d1f90b192f8453370c42fec
+size 10396508

data/train/shard-00038.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6a10ddd0e136d035e4b8d3737eb88557893ff17cb10a6247261a4661537fccf
+size 10225300

data/train/shard-00039.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:683b9bdc6cca3fc103778dce5a4482bc780d903d2295a83ed6e1f065d9404fc8
+size 10321585

data/train/shard-00040.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ed5cf3ba09a7b1d2f8eb3780485c7738e77817867e638837c8ab8a2d534cf9a
+size 10154680

data/train/shard-00041.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45e506540fe40666b3ece2d3658532efda5bdddbe9a34a7a11bfea406948c600
+size 10052079

data/train/shard-00042.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e80612fb96cd6991083d707c7425e582251de4a5c1a7b8e256fcf426756eebf
+size 10250237

data/train/shard-00043.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3112bddf6e5da2a4f4955185c91fa84b8fa6c76213d1347d773aa672d4da5017
+size 9998201

data/train/shard-00044.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:594673319774a18c07b3f08b5ff4a6722e72dc368cfabdcf0e8af185acfc6f07
+size 10494509

data/train/shard-00045.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a573fc93256b4b97d67c17d602538edd05d8b8193b9b70d8c24035732246d537
+size 9913773

data/train/shard-00046.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:063365e63902529599ae1cfe56decbbcec2b2f720fc70aeae54b7903ecf924af
+size 10255108

data/train/shard-00047.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bb4964bf115ee4311a252f087c6e6648cbedb064b31360495d33b6dbd98dfc30
+size 10357882

data/train/shard-00048.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ade20e06f839e17caee89f209066bb59fcbb4af67346f6a82080b8c35b1e6ff0
+size 10011711