Entries Tagged "academic papers"

Page 4 of 76

Undetectable Backdoors in Machine-Learning Models

New paper: “Planting Undetectable Backdoors in Machine Learning Models“:

Abstract: Given the computational cost and technical expertise required to train machine learning models, users may delegate the task of learning to a service provider. We show how a malicious learner can plant an undetectable backdoor into a classifier. On the surface, such a backdoored classifier behaves normally, but in reality, the learner maintains a mechanism for changing the classification of any input, with only a slight perturbation. Importantly, without the appropriate “backdoor key”, the mechanism is hidden and cannot be detected by any computationally-bounded observer. We demonstrate two frameworks for planting undetectable backdoors, with incomparable guarantees.

First, we show how to plant a backdoor in any model, using digital signature schemes. The construction guarantees that given black-box access to the original model and the backdoored version, it is computationally infeasible to find even a single input where they differ. This property implies that the backdoored model has generalization error comparable with the original model. Second, we demonstrate how to insert undetectable backdoors in models trained using the Random Fourier Features (RFF) learning paradigm or in Random ReLU networks. In this construction, undetectability holds against powerful white-box distinguishers: given a complete description of the network and the training data, no efficient distinguisher can guess whether the model is “clean” or contains a backdoor.

Our construction of undetectable backdoors also sheds light on the related issue of robustness to adversarial examples. In particular, our construction can produce a classifier that is indistinguishable from an “adversarially robust” classifier, but where every input has an adversarial example! In summary, the existence of undetectable backdoors represent a significant theoretical roadblock to certifying adversarial robustness.

EDITED TO ADD (4/20): Cory Doctorow wrote about this as well.

Posted on April 19, 2022 at 3:12 PMView Comments

Friday Squid Blogging: Squid Migration and Climate Change

New research on the changing migration of the Doryteuthis opalescens as a result of climate change.

News article:

Stanford researchers have solved a mystery about why a species of squid native to California has been found thriving in the Gulf of Alaska about 1,800 miles north of its expected range: climate change.

As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered.

Read my blog posting guidelines here.

Posted on April 1, 2022 at 4:06 PMView Comments

Friday Squid Blogging: Unexpectedly Low Squid Population in the Arctic

Research:

Abstract: The retreating ice cover of the Central Arctic Ocean (CAO) fuels speculations on future fisheries. However, very little is known about the existence of harvestable fish stocks in this 3.3 million­–square kilometer ecosystem around the North Pole. Crossing the Eurasian Basin, we documented an uninterrupted 3170-kilometer-long deep scattering layer (DSL) with zooplankton and small fish in the Atlantic water layer at 100- to 500-meter depth. Diel vertical migration of this central Arctic DSL was lacking most of the year when daily light variation was absent. Unexpectedly, the DSL also contained low abundances of Atlantic cod, along with lanternfish, armhook squid, and Arctic endemic ice cod. The Atlantic cod originated from Norwegian spawning grounds and had lived in Arctic water temperature for up to 6 years. The potential fish abundance was far below commercially sustainable levels and is expected to remain so because of the low productivity of the CAO.

As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered.

Read my blog posting guidelines here.

Posted on March 25, 2022 at 4:07 PMView Comments

Friday Squid Blog: 328-million-year-old Vampire Squid Ancestor Discovered

A fossilized ancestor of the vampire squid—with ten arms—was discovered and named Syllipsimopodi bideni after President Biden.

Here’s the research paper. Note: Vampire squids are not squids. (Yes, it’s weird.)

As usual, you can also use this squid post to talk about the security stories in the news that I haven’t covered.

Read my blog posting guidelines here.

Posted on March 11, 2022 at 4:01 PMView Comments

Hacking Alexa through Alexa’s Speech

An Alexa can respond to voice commands it issues. This can be exploited:

The attack works by using the device’s speaker to issue voice commands. As long as the speech contains the device wake word (usually “Alexa” or “Echo”) followed by a permissible command, the Echo will carry it out, researchers from Royal Holloway University in London and Italy’s University of Catania found. Even when devices require verbal confirmation before executing sensitive commands, it’s trivial to bypass the measure by adding the word “yes” about six seconds after issuing the command. Attackers can also exploit what the researchers call the “FVV,” or full voice vulnerability, which allows Echos to make self-issued commands without temporarily reducing the device volume.

It does require proximate access, though, at least to set the attack up:

It requires only a few seconds of proximity to a vulnerable device while it’s turned on so an attacker can utter a voice command instructing it to pair with an attacker’s Bluetooth-enabled device. As long as the device remains within radio range of the Echo, the attacker will be able to issue commands.

Research paper.

Posted on March 7, 2022 at 6:20 AMView Comments

Samsung Encryption Flaw

Researchers have found a major encryption flaw in 100 million Samsung Galaxy phones.

From the abstract:

In this work, we expose the cryptographic design and implementation of Android’s Hardware-Backed Keystore in Samsung’s Galaxy S8, S9, S10, S20, and S21 flagship devices. We reversed-engineered and provide a detailed description of the cryptographic design and code structure, and we unveil severe design flaws. We present an IV reuse attack on AES-GCM that allows an attacker to extract hardware-protected key material, and a downgrade attack that makes even the latest Samsung devices vulnerable to the IV reuse attack. We demonstrate working key extraction attacks on the latest devices. We also show the implications of our attacks on two higher-level cryptographic protocols between the TrustZone and a remote server: we demonstrate a working FIDO2 WebAuthn login bypass and a compromise of Google’s Secure Key Import.

Here are the details:

As we discussed in Section 3, the wrapping key used to encrypt the key blobs (HDK) is derived using a salt value computed by the Keymaster TA. In v15 and v20-s9 blobs, the salt is a deterministic function that depends only on the application ID and application data (and constant strings), which the Normal World client fully controls. This means that for a given application, all key blobs will be encrypted using the same key. As the blobs are encrypted in AES-GCM mode-of-operation, the security of the resulting encryption scheme depends on its IV values never being reused.

Gadzooks. That’s a really embarrassing mistake. GSM needs a new nonce for every encryption. Samsung took a secure cipher mode and implemented it insecurely.

News article.

Posted on March 4, 2022 at 6:19 AMView Comments

Decrypting Hive Ransomware Data

Nice piece of research:

Abstract: Among the many types of malicious codes, ransomware poses a major threat. Ransomware encrypts data and demands a ransom in exchange for decryption. As data recovery is impossible if the encryption key is not obtained, some companies suffer from considerable damage, such as the payment of huge amounts of money or the loss of important data. In this paper, we analyzed Hive ransomware, which appeared in June 2021. Hive ransomware has caused immense harm, leading the FBI to issue an alert about it. To minimize the damage caused by Hive Ransomware and to help victims recover their files, we analyzed Hive Ransomware and studied recovery methods. By analyzing the encryption process of Hive ransomware, we confirmed that vulnerabilities exist by using their own encryption algorithm. We have recovered the master key for generating the file encryption key partially, to enable the decryption of data encrypted by Hive ransomware. We recovered 95% of the master key without the attacker’s RSA private key and decrypted the actual infected data. To the best of our knowledge, this is the first successful attempt at decrypting Hive ransomware. It is expected that our method can be used to reduce the damage caused by Hive ransomware.

Here’s the flaw:

The cryptographic vulnerability identified by the researchers concerns the mechanism by which the master keys are generated and stored, with the ransomware strain only encrypting select portions of the file as opposed to the entire contents using two keystreams derived from the master key.

The encryption keystream, which is created from an XOR operation of the two keystreams, is then XORed with the data in alternate blocks to generate the encrypted file. But this technique also makes it possible to guess the keystreams and restore the master key, in turn enabling the decode of encrypted files sans the attacker’s private key.

The researchers said that they were able to weaponize the flaw to devise a method to reliably recover more than 95% of the keys employed during encryption.

Posted on March 1, 2022 at 6:06 AMView Comments

Breaking 256-bit Elliptic Curve Encryption with a Quantum Computer

Researchers have calculated the quantum computer size necessary to break 256-bit elliptic curve public-key cryptography:

Finally, we calculate the number of physical qubits required to break the 256-bit elliptic curve encryption of keys in the Bitcoin network within the small available time frame in which it would actually pose a threat to do so. It would require 317 × 106 physical qubits to break the encryption within one hour using the surface code, a code cycle time of 1 μs, a reaction time of 10 μs, and a physical gate error of 10-3. To instead break the encryption within one day, it would require 13 × 106 physical qubits.

In other words: no time soon. Not even remotely soon. IBM’s largest ever superconducting quantum computer is 127 physical qubits.

Posted on February 9, 2022 at 6:25 AMView Comments

Sidebar photo of Bruce Schneier by Joe MacInnis.