Xiao Wang

28 papers A* 9Journal 18Unranked 1
YearRankTypeTitle / Venue / Authors
2025 J jnl
CoRR
Xiao Wang, Ibrahim Alabdulmohsin, Daniel Salz, Zhe Li, Keran Rong, Xiaohua Zhai
2025 J jnl
CoRR
Michael Tschannen, Alexey A. Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, Olivier J. Hénaff, Jeremiah Harmsen, Andreas Steiner, Xiaohua Zhai
2025 J jnl
CoRR
Brian Gordon, Yonatan Bitton, Andreea Marzoca, Yasumasa Onoe, Xiao Wang, Daniel Cohen-Or, Idan Szpektor
2025 J jnl
CoRR
Jingyun Yang, Ruoyan Avery Yin, Chi Jiang, Yuepeng Hu, Xiaokai Zhu, Xingjian Hu, Sutharsika Kumar, Xiao Wang, Xiaohua Zhai, Keran Rong, Yunyue Zhu, Tianyi Zhang, Zongyou Yin, Jing Kong, Neil Zhenqiang Gong, Zhichu Ren, Haozhe Wang
2024 A* conf
ICLR
Ibrahim Alabdulmohsin, Xiao Wang, Andreas Peter Steiner, Priya Goyal, Alexander D'Amour, Xiaohua Zhai
2024 J jnl
CoRR
Ibrahim Alabdulmohsin, Xiao Wang, Andreas Steiner, Priya Goyal, Alexander D'Amour, Xiaohua Zhai
2024 A* conf
NeurIPS
Bo Wan, Michael Tschannen, Yongqin Xian, Filip Pavetic, Ibrahim M. Alabdulmohsin, Xiao Wang, André Susano Pinto, Andreas Steiner, Lucas Beyer, Xiaohua Zhai
2024 J jnl
CoRR
Bo Wan, Michael Tschannen, Yongqin Xian, Filip Pavetic, Ibrahim Alabdulmohsin, Xiao Wang, André Susano Pinto, Andreas Steiner, Lucas Beyer, Xiaohua Zhai
2024 A* conf
NeurIPS
Angéline Pouget, Lucas Beyer, Emanuele Bugliarello, Xiao Wang, Andreas Steiner, Xiaohua Zhai, Ibrahim M. Alabdulmohsin
2024 J jnl
CoRR
Angéline Pouget, Lucas Beyer, Emanuele Bugliarello, Xiao Wang, Andreas Peter Steiner, Xiaohua Zhai, Ibrahim Alabdulmohsin
2024 A* conf
CVPR
Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, A. J. Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, Radu Soricut
2024 J jnl
CoRR
Andreas Steiner, André Susano Pinto, Michael Tschannen, Daniel Keysers, Xiao Wang, Yonatan Bitton, Alexey A. Gritsenko, Matthias Minderer, Anthony Sherbondy, Shangbang Long, Siyang Qin, R. Reeve Ingle, Emanuele Bugliarello, Sahar Kazemzadeh, Thomas Mesnard, Ibrahim Alabdulmohsin, Lucas Beyer, Xiaohua Zhai
2024 J jnl
CoRR
Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey A. Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bauer, Matko Bosnjak, Xi Chen, Matthias Minderer, Paul Voigtlaender, Ioana Bica, Ivana Balazevic, Joan Puigcerver, Pinelopi Papalampidi, Olivier J. Hénaff, Xi Xiong, Radu Soricut, Jeremiah Harmsen, Xiaohua Zhai
2023 J jnl
CoRR
Lucas Beyer, Bo Wan, Gagan Madan, Filip Pavetic, Andreas Steiner, Alexander Kolesnikov, André Susano Pinto, Emanuele Bugliarello, Xiao Wang, Qihang Yu, Liang-Chieh Chen, Xiaohua Zhai
2023 J jnl
CoRR
Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, Daniel Salz, Xi Xiong, Daniel Vlasic, Filip Pavetic, Keran Rong, Tianli Yu, Daniel Keysers, Xiaohua Zhai, Radu Soricut
2023 J jnl
CoRR
Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, Siamak Shakeri, Mostafa Dehghani, Daniel Salz, Mario Lucic, Michael Tschannen, Arsha Nagrani, Hexiang Hu, Mandar Joshi, Bo Pang, Ceslee Montgomery, Paulina Pietrzyk, Marvin Ritter, A. J. Piergiovanni, Matthias Minderer, Filip Pavetic, Austin Waters, Gang Li, Ibrahim Alabdulmohsin, Lucas Beyer, Julien Amelot, Kenton Lee, Andreas Peter Steiner, Yang Li, Daniel Keysers, Anurag Arnab, Yuanzhong Xu, Keran Rong, Alexander Kolesnikov, Mojtaba Seyedhosseini, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, Radu Soricut
2023 A* conf
ICLR
Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo
2023 A* conf
ICML
Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme Ruiz, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin Fathy Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Collier, Alexey A. Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah J. Harmsen, Neil Houlsby
2023 J jnl
CoRR
Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdulmohsin, Rodolphe Jenatton, Lucas Beyer, Michael Tschannen, Anurag Arnab, Xiao Wang, Carlos Riquelme, Matthias Minderer, Joan Puigcerver, Utku Evci, Manoj Kumar, Sjoerd van Steenkiste, Gamaleldin F. Elsayed, Aravindh Mahendran, Fisher Yu, Avital Oliver, Fantine Huot, Jasmijn Bastings, Mark Patrick Collier, Alexey A. Gritsenko, Vighnesh Birodkar, Cristina Nader Vasconcelos, Yi Tay, Thomas Mensink, Alexander Kolesnikov, Filip Pavetic, Dustin Tran, Thomas Kipf, Mario Lucic, Xiaohua Zhai, Daniel Keysers, Jeremiah Harmsen, Neil Houlsby
2023 A* conf
NeurIPS
Jannik Kossen, Mark Collier, Basil Mustafa, Xiao Wang, Xiaohua Zhai, Lucas Beyer, Andreas Steiner, Jesse Berent, Rodolphe Jenatton, Effrosyni Kokiopoulou
2023 J jnl
CoRR
Jannik Kossen, Mark Collier, Basil Mustafa, Xiao Wang, Xiaohua Zhai, Lucas Beyer, Andreas Steiner, Jesse Berent, Rodolphe Jenatton, Efi Kokiopoulou
2022 A* conf
CVPR
Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, Lucas Beyer
2022 J jnl
CoRR
Xi Chen, Xiao Wang, Soravit Changpinyo, A. J. Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, Alexander Kolesnikov, Joan Puigcerver, Nan Ding, Keran Rong, Hassan Akbari, Gaurav Mishra, Linting Xue, Ashish V. Thapliyal, James Bradbury, Weicheng Kuo, Mojtaba Seyedhosseini, Chao Jia, Burcu Karagol Ayan, Carlos Riquelme, Andreas Steiner, Anelia Angelova, Xiaohua Zhai, Neil Houlsby, Radu Soricut
2022 J jnl
CoRR
Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, Neil Houlsby
2022 conf
ECCV (10)
Matthias Minderer, Alexey A. Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, Neil Houlsby
2021 J jnl
CoRR
Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, Lucas Beyer
2020 A* conf
ICLR
Daniel Keysers, Nathanael Schärli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, Olivier Bousquet
2019 J jnl
CoRR
Daniel Keysers, Nathanael Schärli, Nathan Scales, Hylke Buisman, Daniel Furrer, Sergii Kashubin, Nikola Momchev, Danila Sinopalnikov, Lukasz Stafiniak, Tibor Tihon, Dmitry Tsarkov, Xiao Wang, Marc van Zee, Olivier Bousquet
redb/extractors/apk_extractors/apk_resources.py
← Index redb/extractors/apk_extractors/apk_resources.py python
import hashlib
import inspect
import os
from datetime import datetime, timezone
from typing import Any

from redb.extractors.enum import Tag
from redb.extractors.apk_extractor import APKExtractor
from redb.models.dataclasses import APKResource


# ─── Suspicious file types ──────────────────────────────────────────────
# File types that are suspicious when found inside res/ or assets/.
# Excludes javascript/html (extremely common in legitimate hybrid apps)
# and common media/font types that are normal APK content.
SUSPICIOUS_TYPES = {
    # Executables — no legitimate reason in assets/res
    "elf", "pebin", "macho", "dex", "apk",
    # Java containers — DexClassLoader target
    "jar",
    # Archives — rare in legitimate assets (~135:1 malware-to-benign ratio)
    "zip", "gzip", "7z", "xz", "tar", "bzip2", "rar", "7zip", "lzma",
    # Scripts with system execution capability
    "shell", "python", "powershell", "batch",
}

# ─── Entropy thresholds ─────────────────────────────────────────────────
# For unrecognized/unknown types: encrypted payloads typically land > 7.0
ENTROPY_HIGH_UNKNOWN = 7.0
# For recognized-but-non-image types: stricter threshold
ENTROPY_EXTREME = 7.85

# ─── Android-specific binary format magic bytes ─────────────────────────
# These formats are common in legitimate APKs but unknown to Magika,
# causing misclassification (e.g., AXML → "gzip", profm → "unknown").
AXML_MAGIC = b'\x03\x00\x08\x00'       # Android Binary XML (compiled res/*.xml)
ARSC_MAGIC = b'\x02\x00\x0c\x00'       # Android compiled resource table
ART_PROF_MAGIC = b'pro\x00'            # ART baseline profile
ART_PROFM_MAGIC = b'prm\x00'           # ART baseline profile metadata

# ─── Allowlisted paths ──────────────────────────────────────────────────
# Fixed, hardcoded paths in the Android build system that are always benign.
# ART profiles at these exact paths are shipped by Jetpack ProfileInstaller.
ALLOWLISTED_PATHS = {
    "assets/dexopt/baseline.prof",
    "assets/dexopt/baseline.profm",
}

# ─── Image handling ─────────────────────────────────────────────────────
# Magika-confirmed image types: high entropy is expected (lossy codecs
# like VP8/JPEG arithmetic-code toward entropy ~7.95-8.0 by design).
IMAGE_MAGIKA_TYPES = {"png", "webp", "jpeg", "gif", "bmp", "tiff", "ico"}
IMAGE_EXTENSIONS = {".png", ".webp", ".jpg", ".jpeg", ".gif", ".bmp", ".tiff", ".ico"}

# ─── Types Magika assigns when it can't identify the content ────────────
UNRECOGNIZED_MAGIKA_TYPES = {"unknown", "empty"}

# ─── Resource scan limits ───────────────────────────────────────────────
MAX_RESOURCE_FILES = 5000


class APKResourceExtractor(APKExtractor):

    def __init__(
        self, filepath, log, exporters=None, index_prefix=None,
        known_benign=False, known_malicious=False,
        apk=None,
    ):
        super().__init__(
            filepath, log, exporters, index_prefix,
            known_benign, known_malicious, apk,
        )
        self.resources = []
        self.suspicious_files = []
        self.log.debug(inspect.currentframe().f_code.co_name)

    def tag(self):
        return Tag.APK_RESOURCES.value

    # ─── Core classification logic ──────────────────────────────────────

    def _identify_android_format(self, header: bytes) -> str | None:
        """
        Identify Android-specific binary formats that Magika doesn't know.
        Returns a corrected type label, or None to fall through to Magika.
        """
        if len(header) < 4:
            return None

        magic4 = header[:4]

        # Android Binary XML — all res/*.xml in a compiled APK.
        # Magika often misclassifies this as "gzip".
        if magic4 == AXML_MAGIC:
            return "android_binary_xml"

        # Android compiled resource table (resources.arsc chunks)
        if magic4 == ARSC_MAGIC:
            return "android_resource_table"

        # ART baseline profiles — high entropy (zlib inside) but benign.
        # The format is inert (method reference bitmaps/metadata, not
        # executable code) and some build configs place them at varying paths.
        if magic4 == ART_PROF_MAGIC:
            return "android_art_profile"
        if magic4 == ART_PROFM_MAGIC:
            return "android_art_profile_metadata"

        return None

    def _is_suspicious_resource(
        self, path: str, magika_type: str, entropy: float,
        android_type: str | None,
    ) -> bool:
        """
        Determine if a resource file is suspicious.

        Detection layers:
        1. Allowlisted paths → always benign
        2. Android-specific format override → reclassify Magika mislabels
        3. Image extension vs Magika type mismatch → encrypted blob detection
        4. Magika-confirmed images → benign regardless of entropy
        5. Suspicious type match → flag known-dangerous types
        6. High-entropy unknown blobs → likely encrypted payloads
        """

        # ── Layer 1: Allowlisted paths (hardcoded Android build artifacts) ──
        if path in ALLOWLISTED_PATHS:
            return False

        # ── Layer 2: Android-specific format detection ──────────────────────
        # Override Magika's label for formats it doesn't recognize.
        # All Android-specific formats (AXML, ARSC, ART profiles) are
        # legitimate build artifacts — never suspicious.
        if android_type is not None:
            return False

        # ── Layer 3: Image extension / Magika type mismatch ─────────────────
        # If the file extension claims "image" but Magika's content analysis
        # disagrees, this is a strong signal for an encrypted payload with
        # a fake image extension (e.g., ErrorFather's "rbyypivsnw.png").
        ext = os.path.splitext(path)[1].lower()
        if ext in IMAGE_EXTENSIONS and magika_type not in IMAGE_MAGIKA_TYPES:
            # Exception: Magika might label a valid image as "unknown" if
            # the file is very small (< ~16 bytes). Don't flag tiny files.
            if entropy > 5.0:
                return True

        # ── Layer 4: Magika-confirmed images → benign ───────────────────────
        # Lossy codecs (VP8, JPEG) produce entropy up to ~8.0 by design.
        # If Magika confirms image structure, high entropy is expected.
        if magika_type in IMAGE_MAGIKA_TYPES:
            return False

        # ── Layer 5: Known suspicious file types ────────────────────────────
        if magika_type in SUSPICIOUS_TYPES:
            return True

        # ── Layer 6: High-entropy unrecognized blobs ────────────────────────
        # Files Magika can't identify with high entropy are likely encrypted
        # payloads. Most Android malware packers store encrypted DEX/SO
        # payloads as opaque blobs with random names and no valid magic.
        if magika_type in UNRECOGNIZED_MAGIKA_TYPES and entropy > ENTROPY_HIGH_UNKNOWN:
            return True

        # ── Layer 7: Extreme entropy on any non-image recognized type ───────
        # Catches edge cases where Magika assigns a benign label (e.g.,
        # "xml", "txt") but the entropy is impossibly high for that format.
        if magika_type not in IMAGE_MAGIKA_TYPES and entropy > ENTROPY_EXTREME:
            return True

        return False

    # ─── Extraction pipeline ────────────────────────────────────────────

    def extract(self):
        if not self._is_valid_apk():
            self.log.error(f"Invalid APK for {self.hash.sha256}")
            return None

        try:
            from magika import Magika
            magika = Magika()
        except Exception as e:
            self.log.error(f"Failed to initialize Magika for {self.hash.sha256}: {e}")
            magika = None

        self.resources = []
        self.suspicious_files = []
        scanned = 0

        zf = self._get_zip_file()
        if not zf:
            return None

        with zf:
            for info in zf.infolist():
                if info.is_dir():
                    continue
                if not (info.filename.startswith("res/") or
                        info.filename.startswith("assets/")):
                    continue

                if scanned >= MAX_RESOURCE_FILES:
                    self.log.warning(
                        f"Resource scan limit reached ({MAX_RESOURCE_FILES}), "
                        f"stopping resource enumeration"
                    )
                    break
                scanned += 1

                try:
                    data = zf.read(info.filename)
                except Exception as e:
                    self.log.warning(
                        f"Error reading resource {info.filename}: {e}"
                    )
                    continue

                try:
                    file_sha256 = hashlib.sha256(data).hexdigest()
                    file_entropy = round(self.calculate_entropy(data), 3)

                    # Read first bytes for Android-specific format detection
                    header = data[:16] if len(data) >= 16 else data

                    if magika:
                        try:
                            filetype = magika.identify_bytes(data).output.label
                        except Exception:
                            filetype = "unknown"
                    else:
                        filetype = "unknown"

                    # Identify Android-specific formats once, reuse for
                    # both stored type and suspicion classification
                    android_type = self._identify_android_format(header)
                    stored_type = android_type if android_type else filetype

                    suspicious = self._is_suspicious_resource(
                        path=info.filename,
                        magika_type=filetype,
                        entropy=file_entropy,
                        android_type=android_type,
                    )

                    resource = APKResource(
                        path=info.filename,
                        size=info.file_size,
                        sha256=file_sha256,
                        filetype_magika=stored_type,
                        entropy=file_entropy,
                    )

                    if suspicious:
                        resource.is_suspicious = True
                        self.suspicious_files.append(resource)

                    self.resources.append(resource)
                except Exception as e:
                    self.log.warning(
                        f"Error processing resource {info.filename}: {e}"
                    )
                    continue

        if not self.resources:
            return None

        return {
            "total_resource_count": len(self.resources),
            "total_resource_size": sum(r.size for r in self.resources),
            "suspicious_file_count": len(self.suspicious_files),
            "resources": self.resources,
            "suspicious_files": self.suspicious_files,
        }

    # ─── Export ──────────────────────────────────────────────────────────

    def prepare_export_data(self, exporter_type: str) -> Any:
        if exporter_type == "ClickHouseExporter":
            if not self.resources:
                return None

            current_time = datetime.now(timezone.utc)
            data = []
            for res in self.resources:
                data.append([
                    self.sha256,
                    res.path,
                    res.size,
                    res.sha256,
                    res.filetype_magika,
                    res.entropy,
                    int(res.is_suspicious),
                    current_time,
                ])

            column_names = [
                'sha256', 'resource_path', 'resource_size',
                'resource_sha256', 'resource_magika', 'resource_entropy',
                'is_suspicious', 'analysis_date',
            ]

            column_type_names = [
                'FixedString(64)', 'String', 'UInt64',
                'FixedString(64)', 'LowCardinality(String)', 'Float32',
                'UInt8', "DateTime64(3, 'UTC')",
            ]

            return (data, column_names, column_type_names)

    def get_clickhouse_table(self) -> str:
        return "redb_apk_resources"