James T. Teo

37 papers Journal 29Unranked 8
YearRankTypeTitle / Venue / Authors
2026 J jnl
CoRR
Yujian Gan, Stephen H. Barlow, Ben Holgate, Joe Davies, James T. Teo, Joel S. Winston, Mark P. Richardson
2025 conf
ACL (6)
Shubham Agarwal, Thomas Searle, Mart Ratas, Anthony Shek, James T. Teo, Richard J. B. Dobson
2025 J jnl
Comput. Methods Programs Biomed.
Henry C. Watkins, Robert J. Gray, Adam Julius, Yee-Haur Mah, James T. Teo, Walter H. L. Pinaya, Paul Wright, Ashwani Jha, Holger Engleitner, Jorge Cardoso, Sébastien Ourselin, Geraint Rees, Hans Rolf Jäger, Parashkev Nachev
2025 J jnl
CoRR
Shubham Agarwal, Vlad Dinu, Thomas Searle, Mart Ratas, Anthony Shek, Dan F. Stein, James T. Teo, Richard J. B. Dobson
2025 conf
ACL (6)
Nicole Ren, James T. Teo
2025 J jnl
npj Digit. Medicine
Hugh Logan Ellis, Edward Palmer, James T. Teo, Martin Whyte, Kenneth Rockwood, Zina M. Ibrahim
2024 J jnl
BMC Medical Informatics Decis. Mak.
Mohammad Al-Agil, Stephen J. Obee, Vlad Dinu, James T. Teo, David Brawand, Piers E. M. Patten, Anwar Alhaq
2024 conf
BioNLP@ACL
Ben Holgate, Shichao Fang, Anthony Shek, Matthew McWilliam, Pedro F. Viana, Joel S. Winston, James T. Teo, Mark P. Richardson
2024 J jnl
CoRR
Chayanin Tangwiriyasakul, Pedro Borges, Stefano Moriconi, Paul Wright, Yee-Haur Mah, James T. Teo, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2024 J jnl
CoRR
Shubham Agarwal, Thomas Searle, Mart Ratas, Anthony Shek, James T. Teo, Richard J. B. Dobson
2024 J jnl
CoRR
Zeljko Kraljevic, Joshua Au Yeung, Daniel Bean, James T. Teo, Richard J. B. Dobson
2024 J jnl
BMC Medical Informatics Decis. Mak.
Joshua Au Yeung, Anthony Shek, Thomas Searle, Zeljko Kraljevic, Vlad Dinu, Mart Ratas, Mohammad Al-Agil, Aleksandra Foy, Barbara Rafferty, Vitaliy Oliynyk, James T. Teo
2023 J jnl
Frontiers Digit. Health
Joshua Au Yeung, Zeljko Kraljevic, Akish Luintel, Alfred Balston, Esther Idowu, Richard J. B. Dobson, James T. Teo
2023 J jnl
J. Biomed. Informatics
Thomas Searle, Zina M. Ibrahim, James T. Teo, Richard J. B. Dobson
2023 conf
BrainLes/SWITCH@MICCAI
Chayanin Tangwiriyasakul, Pedro Borges, Stefano Moriconi, Paul Wright, Yee-Haur Mah, James T. Teo, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2023 J jnl
Medical Image Anal.
Mark S. Graham, Petru-Daniel Tudosiu, Paul Wright, Walter Hugo Lopez Pinaya, Petteri Teikari, Ashay Patel, Jean-Marie U.-King-Im, Yee-Haur Mah, James T. Teo, Hans Rolf Jäger, David Werring, Geraint Rees, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2023 conf
MICCAI (1)
Mark S. Graham, Walter Hugo Lopez Pinaya, Paul Wright, Petru-Daniel Tudosiu, Yee-Haur Mah, James T. Teo, Hans Rolf Jäger, David Werring, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2023 J jnl
CoRR
Mark S. Graham, Walter Hugo Lopez Pinaya, Paul Wright, Petru-Daniel Tudosiu, Yee-Haur Mah, James T. Teo, Hans Rolf Jäger, David Werring, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2023 conf
ICHI
Zeljko Kraljevic, Anthony Shek, Joshua Au Yeung, Ewart Jonathan Sheldon, Haris Shuaib, Mohammad Al-Agil, Xi Bai, Kawsar Noor, Anoop D. Shah, Richard J. B. Dobson, James T. Teo
2023 J jnl
CoRR
Zeljko Kraljevic, Anthony Shek, Joshua Au Yeung, Ewart Jonathan Sheldon, Mohammad Al-Agil, Haris Shuaib, Xi Bai, Kawsar Noor, Anoop D. Shah, Richard J. B. Dobson, James T. Teo
2022 J jnl
IEEE J. Biomed. Health Informatics
Zina M. Ibrahim, Daniel Bean, Thomas Searle, Linglong Qian, Honghan Wu, Anthony Shek, Zeljko Kraljevic, James Galloway, Sam Norton, James T. Teo, Richard J. B. Dobson
2022 conf
MICCAI (8)
Walter H. L. Pinaya, Mark S. Graham, Robert J. Gray, Pedro F. Da Costa, Petru-Daniel Tudosiu, Paul Wright, Yee-Haur Mah, Andrew D. MacKinnon, James T. Teo, Hans Rolf Jäger, David Werring, Geraint Rees, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2022 J jnl
CoRR
Walter H. L. Pinaya, Mark S. Graham, Robert J. Gray, Pedro F. Da Costa, Petru-Daniel Tudosiu, Paul Wright, Yee-Haur Mah, Andrew D. MacKinnon, James T. Teo, Hans Rolf Jäger, David Werring, Geraint Rees, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2022 J jnl
CoRR
Zeljko Kraljevic, Daniel Bean, Anthony Shek, Rebecca Bendayan, Joshua Au Yeung, Alexander Deng, Alfie Baston, Jack Ross, Esther Idowu, James T. Teo, Richard J. B. Dobson
2022 J jnl
npj Digit. Medicine
Joe Zhang, Sanjay Budhdeo, William Wasswa, Paul Cerrato, Haris Shuaib, Harpreet Sood, Hutan Ashrafian, John D. Halamka, James T. Teo
2022 J jnl
CoRR
Thomas Searle, Zina M. Ibrahim, James T. Teo, Richard J. B. Dobson
2022 conf
MIDL
Mark S. Graham, Petru-Daniel Tudosiu, Paul Wright, Walter Hugo Lopez Pinaya, Jean-Marie U.-King-Im, Yee-Haur Mah, James T. Teo, Hans Rolf Jäger, David Werring, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2022 J jnl
CoRR
Mark S. Graham, Petru-Daniel Tudosiu, Paul Wright, Walter Hugo Lopez Pinaya, Jean-Marie U.-King-Im, Yee-Haur Mah, James T. Teo, Hans Rolf Jäger, David Werring, Parashkev Nachev, Sébastien Ourselin, M. Jorge Cardoso
2021 J jnl
J. Am. Medical Informatics Assoc.
Honghan Wu, Huayu Zhang, Andreas Karwath, Zina M. Ibrahim, Ting Shi, Xin Zhang, Kun Wang, Jiaxing Sun, Kevin Dhaliwal, Daniel Bean, Victor Roth Cardoso, Kezhi Li, James T. Teo, Amitava Banerjee, Fang Gao-Smith, Tony Whitehouse, Tonny Veenith, Georgios V. Gkoutos, Xiaodong Wu, Richard J. B. Dobson, Bruce Guthrie
2021 J jnl
CoRR
Thomas Searle, Zina M. Ibrahim, James T. Teo, Richard J. B. Dobson
2021 J jnl
J. Biomed. Informatics
Thomas Searle, Zina M. Ibrahim, James T. Teo, Richard J. B. Dobson
2021 J jnl
CoRR
Zeljko Kraljevic, Anthony Shek, Daniel Bean, Rebecca Bendayan, James T. Teo, Richard J. B. Dobson
2021 J jnl
Artif. Intell. Medicine
Zeljko Kraljevic, Thomas Searle, Anthony Shek, Lukasz Roguski, Kawsar Noor, Daniel Bean, Aurelie Mascio, Leilei Zhu, Amos A. Folarin, Angus Roberts, Rebecca Bendayan, Mark P. Richardson, Robert Stewart, Anoop D. Shah, Wai Keong Wong, Zina M. Ibrahim, James T. Teo, Richard J. B. Dobson
2021 J jnl
npj Digit. Medicine
James T. Teo, Vlad Dinu, William Bernal, Phil Davidson, Vitaliy Oliynyk, Cormac Breen, Richard D. Barker, Richard J. B. Dobson
2020 J jnl
CoRR
Zina M. Ibrahim, Daniel Bean, Thomas Searle, Honghan Wu, Anthony Shek, Zeljko Kraljevic, James Galloway, Sam Norton, James T. Teo, Richard J. B. Dobson
2020 J jnl
CoRR
Zeljko Kraljevic, Thomas Searle, Anthony Shek, Lukasz Roguski, Kawsar Noor, Daniel Bean, Aurelie Mascio, Leilei Zhu, Amos A. Folarin, Angus Roberts, Rebecca Bendayan, Mark P. Richardson, Robert Stewart, Anoop D. Shah, Wai Keong Wong, Zina M. Ibrahim, James T. Teo, Richard J. B. Dobson
2010 J jnl
J. Cogn. Neurosci.
Leonora Wilkinson, James T. Teo, Ignacio Obeso, John C. Rothwell, Marjan Jahanshahi
redb/extractors/decompiler/apk/analyzer.py
← Index redb/extractors/decompiler/apk/analyzer.py python
"""APK Code Analyzer — orchestrates androguard + JADX + apktool analysis.

This is the APK equivalent of BinaryNinjaDecompiler in
redb/extractors/decompiler/bninja/decompiler.py.
"""

import math
import os
import tempfile
from collections import Counter
from concurrent.futures import ThreadPoolExecutor, as_completed
from typing import Any, Dict, List, Optional, Set

from redb.extractors.decompiler.apk.apktool_wrapper import ApktoolDisassembler
from redb.extractors.decompiler.apk.jadx_wrapper import JADXDecompiler
from redb.extractors.decompiler.apk.library_filter import LibraryFilter
from redb.extractors.decompiler.apk.method_extractor import (
    compute_minhash,
    compute_prime_product_smali,
    compute_sha256,
    compute_ssdeep,
    compute_tlsh,
    count_call_instructions,
    dalvik_to_java_class,
    dalvik_to_java_prototype,
    detect_obfuscation_indicators,
)
from redb.extractors.decompiler.apk.smali_normalization import normalize_method_body
from redb.extractors.decompiler.apk.smali_cfg import compute_cfg_metrics
from redb.extractors.decompiler.apk.smali_parser import SmaliParser


class APKCodeAnalyzer:
    """Orchestrates APK code analysis combining three tools.

    Combines androguard (call graphs, xrefs, method enumeration),
    JADX (Java decompilation), and apktool (smali disassembly).
    """

    def __init__(
        self,
        filepath: str,
        timeout: int = 600,
        log=None,
        decompile_modules: Set[str] = None,
    ):
        self.filepath = filepath
        self.timeout = timeout
        self.log = log
        self.decompile_modules = decompile_modules or {"all"}
        self.min_instructions = int(
            os.getenv("APK_MIN_METHOD_INSTRUCTIONS", "5")
        )
        self.library_filter = LibraryFilter()
        self.jadx = JADXDecompiler(log=log)
        self.apktool = ApktoolDisassembler(log=log)

        self._temp_dirs = []

    def extract(self) -> Dict[str, Any]:
        """Run full APK code analysis.

        Returns dict with keys:
            decompiled_content, decompiled_refs,
            smali_content, smali_refs,
            similarity_metrics, strings, analysis_errors
        """
        results = {
            "decompiled_content": [],
            "decompiled_refs": [],
            "smali_content": [],
            "smali_refs": [],
            "similarity_metrics": [],
            "cfg": [],
            "strings": [],
            "analysis_errors": [],
        }

        # Create temp directories
        jadx_dir = tempfile.mkdtemp(prefix="redb_jadx_")
        apktool_dir = tempfile.mkdtemp(prefix="redb_apktool_")
        self._temp_dirs.extend([jadx_dir, apktool_dir])

        # Step 1-3: Run JADX, apktool, and androguard in parallel
        jadx_success = False
        apktool_success = False
        androguard_result = None

        with ThreadPoolExecutor(max_workers=3) as executor:
            futures = {}
            futures[executor.submit(self.jadx.decompile, self.filepath, jadx_dir)] = "jadx"
            futures[executor.submit(self.apktool.disassemble, self.filepath, apktool_dir)] = "apktool"
            futures[executor.submit(self._run_androguard)] = "androguard"

            for future in as_completed(futures):
                tool = futures[future]
                try:
                    result = future.result()
                    if tool == "jadx":
                        jadx_success = result
                    elif tool == "apktool":
                        apktool_success = result
                    elif tool == "androguard":
                        androguard_result = result
                except Exception as e:
                    if self.log:
                        self.log.error(f"{tool} failed: {e}")
                    results["analysis_errors"].append({
                        "class_name": None,
                        "method_name": None,
                        "error_location": tool,
                        "error_message": str(e),
                        "error_type": type(e).__name__,
                    })

        if androguard_result is None:
            if self.log:
                self.log.error("Androguard analysis failed — cannot proceed")
            return results

        apk_obj, dexs, analysis = androguard_result

        # Step 4b: Extract strings from DEX files
        try:
            results["strings"] = self._extract_strings(dexs)
        except Exception as e:
            if self.log:
                self.log.error(f"String extraction failed: {e}")
            results["analysis_errors"].append({
                "class_name": None,
                "method_name": None,
                "error_location": "strings",
                "error_message": f"String extraction failed: {e}",
                "error_type": type(e).__name__,
            })

        # Step 5: Parse smali output
        smali_methods = {}
        if apktool_success:
            try:
                smali_dirs = self.apktool.get_smali_directories(apktool_dir)
                for sdir in smali_dirs:
                    smali_methods.update(SmaliParser.parse_smali_directory(sdir))
            except Exception as e:
                if self.log:
                    self.log.error(f"Smali parsing failed: {e}")
                results["analysis_errors"].append({
                    "class_name": None,
                    "method_name": None,
                    "error_location": "apktool",
                    "error_message": f"Smali parsing failed: {e}",
                    "error_type": type(e).__name__,
                })

        # Step 6: Parse Java output
        java_methods = {}
        if jadx_success:
            try:
                java_methods = self.jadx.parse_java_methods(jadx_dir)
            except Exception as e:
                if self.log:
                    self.log.error(f"Java parsing failed: {e}")
                results["analysis_errors"].append({
                    "class_name": None,
                    "method_name": None,
                    "error_location": "jadx",
                    "error_message": f"Java parsing failed: {e}",
                    "error_type": type(e).__name__,
                })

        if not apktool_success and self.log:
            self.log.info(
                "apktool failed — falling back to androguard disassembly"
            )

        # Step 7-8: Process each method from androguard
        seen_decompiled_hashes = set()
        seen_smali_hashes = set()

        try:
            for method in analysis.get_methods():
                try:
                    self._process_method(
                        method,
                        smali_methods,
                        java_methods,
                        results,
                        seen_decompiled_hashes,
                        seen_smali_hashes,
                    )
                except Exception as e:
                    method_name = "unknown"
                    class_name = "unknown"
                    try:
                        if not method.is_external():
                            enc = method.get_method()
                            class_name = enc.get_class_name()
                            method_name = enc.get_name()
                    except Exception:
                        pass
                    results["analysis_errors"].append({
                        "class_name": class_name,
                        "method_name": method_name,
                        "error_location": "analysis",
                        "error_message": str(e),
                        "error_type": type(e).__name__,
                    })
        except Exception as e:
            if self.log:
                self.log.error(f"Method enumeration failed: {e}")
            results["analysis_errors"].append({
                "class_name": None,
                "method_name": None,
                "error_location": "androguard",
                "error_message": f"Method enumeration failed: {e}",
                "error_type": type(e).__name__,
            })

        if self.log:
            stats = self.library_filter.get_filter_stats()
            self.log.info(
                f"APK analysis complete: {stats['user']} user methods, "
                f"{stats['library']} library methods filtered, "
                f"{len(results['decompiled_content'])} decompiled, "
                f"{len(results['smali_content'])} smali, "
                f"{len(results['strings'])} strings"
            )

        return results

    def _run_androguard(self):
        """Run androguard analysis on the APK."""
        from androguard.misc import AnalyzeAPK
        return AnalyzeAPK(self.filepath)

    def _process_method(
        self,
        method,
        smali_methods: Dict,
        java_methods: Dict,
        results: Dict,
        seen_decompiled: set,
        seen_smali: set,
    ):
        """Process a single method from androguard analysis."""
        # Skip external methods (no code body)
        if method.is_external():
            return

        encoded = method.get_method()
        class_name = encoded.get_class_name()
        method_name = encoded.get_name()
        descriptor = encoded.get_descriptor()

        # Build method key for cross-tool matching
        method_key = SmaliParser.make_method_key(class_name, method_name, descriptor)

        # Check if library
        is_lib = self.library_filter.is_library(class_name)
        method_type = "LIBRARY" if is_lib else "USER"

        # Skip library methods for content tables (but they're still in xrefs)
        if is_lib:
            return

        # Get xrefs
        # In androguard 4.x, xref tuples are (ClassAnalysis, MethodAnalysis, offset).
        # The MethodAnalysis wrapper doesn't expose get_class_name()/get_name()
        # directly — we need to unwrap via .get_method() first.
        callers = []
        callees = []
        try:
            for ref_class, ref_method, offset in method.get_xref_from():
                try:
                    enc = ref_method.get_method()
                    caller_key = f"{enc.get_class_name()}->{enc.get_name()}"
                except AttributeError:
                    # Fallback for older androguard where ref_method is EncodedMethod
                    caller_key = f"{ref_method.get_class_name()}->{ref_method.get_name()}"
                callers.append(caller_key)
        except Exception:
            pass

        try:
            for ref_class, ref_method, offset in method.get_xref_to():
                try:
                    enc = ref_method.get_method()
                    callee_key = f"{enc.get_class_name()}->{enc.get_name()}"
                except AttributeError:
                    callee_key = f"{ref_method.get_class_name()}->{ref_method.get_name()}"
                callees.append(callee_key)
        except Exception:
            pass

        # Look up smali body
        smali_method = smali_methods.get(method_key)
        smali_body = None
        smali_normalized = None
        smali_hash = None
        instruction_count = 0
        register_count = 0

        if smali_method:
            smali_body = smali_method.body
            instruction_count = smali_method.instruction_count
            register_count = smali_method.register_count
        else:
            # Fallback: use androguard's own disassembler
            smali_body, instruction_count, register_count = (
                self._disassemble_with_androguard(encoded)
            )

        # Apply minimum instruction filter
        if instruction_count < self.min_instructions:
            return

        if smali_body:
            smali_normalized = SmaliParser.normalize_smali_body(smali_body)
            if smali_normalized:
                smali_hash = compute_sha256(smali_normalized)

        if not smali_hash:
            return

        # Compute obfuscation indicators from smali
        obfuscation = detect_obfuscation_indicators(
            method_name, class_name, smali_body, instruction_count
        )

        # Compute CFG metrics early — needed for both content tables and cfg table
        cfg_metrics = None
        if smali_body:
            try:
                cfg_metrics = compute_cfg_metrics(smali_body)
            except Exception:
                pass

        # Smali content (deduplicated)
        if smali_hash and smali_hash not in seen_smali:
            seen_smali.add(smali_hash)
            results["smali_content"].append({
                "smali_method_hash": smali_hash,
                "smali_method": smali_body,
                "smali_method_type": method_type,
                "smali_instructions_count": instruction_count,
                "smali_register_count": register_count,
                "smali_has_string_encryption": obfuscation.get("has_string_encryption", False),
                "smali_has_reflection_calls": obfuscation.get("has_reflection_calls", False),
                "smali_excessive_goto_count": obfuscation.get("excessive_goto_count", False),
                "smali_flattened_score": cfg_metrics.flattened_score if cfg_metrics else 0.0,
                "smali_mba_score": cfg_metrics.mba_score if cfg_metrics else 0.0,
            })

        # Smali reference (per-binary)
        ssdeep_val = compute_ssdeep(smali_normalized) if smali_normalized else None
        tlsh_val = compute_tlsh(smali_normalized) if smali_normalized else None

        # Semantically normalized ssdeep/TLSH (analogous to Binja's ssdeep_llil/tlsh_llil)
        ssdeep_normalized_val = None
        tlsh_normalized_val = None
        if smali_normalized:
            normalized_lines = normalize_method_body(smali_normalized, level="opcode_api")
            if normalized_lines:
                normalized_text = "\n".join(normalized_lines)
                ssdeep_normalized_val = compute_ssdeep(normalized_text)
                tlsh_normalized_val = compute_tlsh(normalized_text)

        # Look up Java source
        java_class = dalvik_to_java_class(class_name)
        java_key_candidates = [
            f"{java_class}.{method_name}",
        ]
        java_source = None
        decompiled_hash = None

        for jk in java_key_candidates:
            for key, source in java_methods.items():
                if key.startswith(jk):
                    java_source = source
                    break
            if java_source:
                break

        if java_source:
            # Normalize and hash decompiled content
            normalized_java = _normalize_java(java_source)
            decompiled_hash = compute_sha256(normalized_java)

            # Decompiled content (deduplicated)
            if decompiled_hash not in seen_decompiled:
                seen_decompiled.add(decompiled_hash)
                results["decompiled_content"].append({
                    "decompiled_method_hash": decompiled_hash,
                    "decompiled_method": java_source,
                    "decompiled_method_type": method_type,
                    "decompiled_has_string_encryption": obfuscation.get("has_string_encryption", False),
                    "decompiled_has_reflection_calls": obfuscation.get("has_reflection_calls", False),
                    "decompiled_excessive_goto_count": obfuscation.get("excessive_goto_count", False),
                })

            # Decompiled reference
            method_prototype = dalvik_to_java_prototype(
                method_name, descriptor, class_name
            )
            results["decompiled_refs"].append({
                "decompiled_method_hash": decompiled_hash,
                "smali_method_hash": smali_hash,
                "decompiled_class_name": dalvik_to_java_class(class_name),
                "decompiled_method_name": method_name,
                "decompiled_method_signature": descriptor,
                "decompiled_method_prototype": method_prototype,
                "functions_caller": callers,
                "functions_call": callees,
            })

        # Smali reference
        if smali_hash:
            results["smali_refs"].append({
                "smali_method_hash": smali_hash,
                "decompiled_method_hash": decompiled_hash,
                "smali_class_name": dalvik_to_java_class(class_name),
                "smali_method_name": method_name,
                "smali_method_signature": descriptor,
                "ssdeep_smali": ssdeep_val,
                "tlsh_smali": tlsh_val,
            })

        # Similarity metrics (content-based fuzzy matching only)
        if smali_hash and smali_normalized:
            minhash_sig = compute_minhash(smali_normalized)
            sim_entry = {
                "smali_method_hash": smali_hash,
                "cyclomatic_complexity": cfg_metrics.cyclomatic_complexity if cfg_metrics else None,
                "ssdeep_smali": ssdeep_val,
                "tlsh_smali": tlsh_val,
                "ssdeep_smali_normalized": ssdeep_normalized_val,
                "tlsh_smali_normalized": tlsh_normalized_val,
                "minhash": minhash_sig or [],
            }
            results["similarity_metrics"].append(sim_entry)

        # CFG entry (structural/topological features)
        if smali_hash and cfg_metrics and smali_body:
            prime_product = compute_prime_product_smali(smali_body)
            call_count = count_call_instructions(smali_body)
            cfg_entry = {
                "smali_method_hash": smali_hash,
                "cfg_topology_hash": cfg_metrics.cfg_topology_hash,
                "block_count": cfg_metrics.block_count,
                "edge_count": cfg_metrics.edge_count,
                "cfg_instructions_count": instruction_count,
                "call_count": call_count,
                "cyclomatic_complexity": cfg_metrics.cyclomatic_complexity,
                "loop_count": cfg_metrics.loop_count,
                "max_depth": cfg_metrics.max_depth,
                "max_fan_out": cfg_metrics.max_fan_out,
                "md_index_topdown": cfg_metrics.md_index_topdown,
                "md_index_bottomup": cfg_metrics.md_index_bottomup,
                "prime_product_smali": prime_product,
                "cfg_feature_tlsh": cfg_metrics.cfg_feature_tlsh,
                "wl_minhash": cfg_metrics.wl_minhash,
                "bb_features": cfg_metrics.block_features,
                "cfg_adjacency": cfg_metrics.cfg_adjacency,
            }
            results["cfg"].append(cfg_entry)

    @staticmethod
    def _string_entropy(s: str) -> float:
        """Compute Shannon entropy of a string."""
        if not s:
            return 0.0
        freq = Counter(s)
        length = len(s)
        return -sum(
            (count / length) * math.log2(count / length)
            for count in freq.values()
        )

    def _extract_strings(self, dexs) -> List[Dict[str, Any]]:
        """Extract deduplicated strings from all DEX objects.

        Uses androguard's get_strings() on each DEX, deduplicates by value,
        and computes entropy — matching the Binja StringAnalysis output format
        so IOCExtractorFromResults can consume them identically.

        Returns list of dicts with keys:
            string, string_encoding, string_offset, string_length, string_entropy
        """
        seen = set()
        strings = []
        offset_counter = 0

        for dex in dexs:
            try:
                dex_strings = dex.get_strings()
            except Exception:
                continue

            if not dex_strings:
                continue

            for s in dex_strings:
                if not s or s in seen:
                    continue
                # DEX MUTF-8 strings may contain unpaired surrogates (e.g. \ud800)
                # that are invalid UTF-8 and will fail ClickHouse insert
                s = s.encode('utf-8', errors='replace').decode('utf-8')
                if not s or s in seen:
                    continue
                seen.add(s)

                strings.append({
                    "string": s,
                    "string_encoding": "UTF8",
                    "string_offset": offset_counter,
                    "string_length": len(s),
                    "string_entropy": self._string_entropy(s),
                })
                offset_counter += 1

        return strings

    @staticmethod
    def _disassemble_with_androguard(encoded):
        """Fallback disassembly using androguard when apktool fails.

        Returns (smali_body, instruction_count, register_count).
        """
        code = encoded.get_code()
        if not code:
            return None, 0, 0

        register_count = code.get_registers_size()
        lines = []
        instruction_count = 0

        try:
            bytecode = code.get_bc()
            if not bytecode:
                return None, 0, register_count

            for instruction in bytecode.get_instructions():
                op_name = instruction.get_name()
                output = instruction.get_output()
                if output:
                    output = _normalize_androguard_operands(op_name, output)
                    lines.append(f"    {op_name} {output}")
                else:
                    lines.append(f"    {op_name}")
                instruction_count += 1
        except Exception:
            return None, 0, register_count

        if not lines:
            return None, 0, register_count

        smali_body = "\n".join(lines)
        return smali_body, instruction_count, register_count

    def cleanup(self):
        """Remove temporary directories."""
        import shutil
        for d in self._temp_dirs:
            try:
                if os.path.isdir(d):
                    shutil.rmtree(d)
            except Exception:
                pass
        self._temp_dirs.clear()


def _normalize_androguard_operands(op_name: str, output: str) -> str:
    """Normalize androguard instruction output to match apktool smali format.

    Androguard's get_output() differs from apktool in two key ways:
    1. invoke-* operands lack {braces} around registers:
       androguard: 'v0, v1, Lcom/Foo;->bar()V'
       apktool:    '{v0, v1}, Lcom/Foo;->bar()V'
    2. Field instructions use a space instead of colon between name and type:
       androguard: 'v0, v1, LA;->field Ljava/lang/String;'
       apktool:    'v0, v1, LA;->field:Ljava/lang/String;'
    """
    # invoke-* instructions: wrap register args in {braces}
    if op_name.startswith("invoke-"):
        # Find the class/method reference (starts with L or [)
        # Split on ', ' and find where the reference begins
        parts = output.split(", ")
        reg_parts = []
        ref_idx = None
        for i, part in enumerate(parts):
            stripped = part.strip()
            if stripped.startswith("L") or stripped.startswith("["):
                ref_idx = i
                break
            reg_parts.append(part)

        if ref_idx is not None and reg_parts:
            regs = ", ".join(reg_parts)
            ref = ", ".join(parts[ref_idx:])
            return "{" + regs + "}, " + ref
        elif reg_parts:
            # No reference found — just wrap all as registers
            return "{" + ", ".join(reg_parts) + "}"

    # Field instructions: fix 'field Ltype;' → 'field:Ltype;'
    if op_name.startswith(("iget", "iput", "sget", "sput")):
        # Pattern: '... ClassName;->fieldName Ltype;' or '... ClassName;->fieldName [Ltype;'
        # The space between fieldName and the type descriptor should be a colon
        arrow_idx = output.find("->")
        if arrow_idx != -1:
            after_arrow = output[arrow_idx + 2:]
            # Find the space before the type descriptor
            space_idx = after_arrow.find(" ")
            if space_idx != -1:
                remaining = after_arrow[space_idx + 1:]
                # Check that what follows is a type descriptor
                if remaining.startswith(("L", "[", "Z", "B", "S", "C",
                                         "I", "J", "F", "D")):
                    after_arrow = after_arrow[:space_idx] + ":" + remaining
                    output = output[:arrow_idx + 2] + after_arrow

    return output


def _normalize_java(source: str) -> str:
    """Normalize Java source for consistent hashing.

    Strip leading/trailing whitespace, normalize indentation.
    """
    lines = []
    for line in source.split("\n"):
        stripped = line.strip()
        if stripped:
            lines.append(stripped)
    return "\n".join(lines)