Sap Se

Sap Ag

Martin Haerterich

Florian Kerschbaum

Axel Schroepfer

Andreas Schaad

Mathias Kohler

Walter Tighzert

Martin Johns

Isabelle Hang

Patrick Grofig

Daniel Bernau

Knut Heusermann

Matthias Becker

Guang Yang

Torsten Buecheler

Olga Kreindlina

Sophie Kraut

Christian Hohmann

Dietmar Henkes

Oliver Berger

Tom Ganz

Xenia Rieger

Walter Zimmermann

Marcus Echter

Albert Neumueller

Thomas Vogt

Stefan J Moeller

Benjamin Weggenmann

Philip-William Grassal

Christoph Haefner

Philipp Rall

Benjamin Hilprecht

Mukesh Kumar

Florian Hahn

Mohit V Gadkari

Santhosh N

Steffen Witt

Kritesh Vasing

Florian Knoerzer

Hannah Keller

Erik Imgrund

Marius Musch

Karthik Narayanan Mohan

Applications may contain vulnerabilities to attack via malicious inputs. Machine-learning models may be trained to detect these vulnerabilities by accepting source code as input and outputting a probability that each of a set of vulnerabilities exists in the source code. Explanation methods may identify one or more locations within the source code that are likely to cause the vulnerability. Directed fuzzing provides a range of inputs to source code. The inputs that cause the source code to fail are detected and the portions of the source code that were vulnerable are identified. The results of the directed fuzzing are used to select between explanations generated by multiple explanation methods, to provide additional training data to a machine-learning model, to provide additional training data to an explanation method, or any suitable combination thereof.

Directed fuzzing for vulnerability detection

A system and method include reception of a plurality of code samples and corresponding ground truth classifications, generation, for each of the plurality code samples, of a plurality of perturbed code samples, generation of first probabilities for each of the plurality of code samples and the perturbed code samples using a first classification model, generation of second probabilities for each of the plurality of code samples and the perturbed code samples using a second model, determination of a causal probability for each code sample based on the first probabilities and the second probabilities, determination of a causal performance metric of the first classification model based on the ground truth classifications and the causal probabilities, and presentation of the causal performance metric.

Measuring confounding effects in machine learning-based vulnerability discovery

Various examples are directed to systems and methods for obscuring private information in input data. A system may apply an encoder model to an input data unit to generate a latent space representation of the input data unit. The system may apply multi-dimensional noise to the latent space representation of the input data unit, the multi-dimensional noise having a first value in a first latent space dimension and a second value different than the first value in a second latent space dimension. The system may apply a decoder model to the latent space representation of the input data unit to generate an obscured data unit.

Data obscuring for privacy-enhancement

Data is received that specifies a bound for an adversarial posterior belief ρthat corresponds to a likelihood to re-identify data points from the dataset based on a differentially private function output. Privacy parameters ε, δ are then calculated based on the received data that govern a differential privacy (DP) algorithm to be applied to a function to be evaluated over a dataset. The calculating is based on a ratio of probabilities distributions of different observations, which are bound by the posterior belief ρas applied to a dataset. The calculated privacy parameters are then used to apply the DP algorithm to the function over the dataset. Related apparatus, systems, techniques and articles are also described.

Interpretability framework for differentially private deep learning

Techniques for implementing a differentially private variational autoencoder for data obfuscation are disclosed. In some embodiments, a computer system performs operations comprising: encoding input data into a latent space representation of the input data, the encoding of the input data comprising: inferring latent space parameters of a latent space distribution based on the input data, the latent space parameters comprising a mean and a standard deviation, the inferring of the latent space parameters comprising bounding the mean within a finite space and using a global value for the standard deviation, the global value being independent of the input data; and sampling data from the latent space distribution; and decoding the sampled data of the latent space representation into output data.

Differentially private variational autoencoders for data obfuscation

A system is described that can include a machine learning model and at least one programmable processor communicatively coupled to the machine learning model. The machine learning model can receive data, generate a continuous probability distribution associated with the data, sample a latent variable from the continuous probability distribution to generate a plurality of samples, and generate reconstructed data from the plurality of samples. The at least one programmable processor can compute a reconstruction error by determining a distance between the reconstructed data and the data, and generate, based on the reconstruction error, an indication representing whether a specific record within the received data was used to train the machine learning model. Related apparatuses, methods, techniques, non-transitory computer programmable products, non-transitory machine-readable medium, articles, and other systems are also within the scope of this disclosure.

Accurately identifying members of training data in variational autoencoders by reconstruction error

Various examples are directed to systems and methods for detecting training data for a generative model. A computer system may access generative model sample data and a first test sample. The computer system may determine whether a first generative model sample of the plurality of generative model samples is within a threshold distance of the first test sample and whether a second generative model sample of the plurality of generative model samples is within the threshold distance of the first test sample. The computer system may determine that a probability that the generative model was trained with the first test sample is greater than or equal to a threshold probability based at least in part on whether the first generative model sample is within the threshold distance of the first test sample, the determining also based at least in part on whether the second generative model sample is within the threshold distance of the first test sample.

Computer systems for detecting training data usage in generative models

A client comprising a web browser is provided. The client is configured to: run an application in the web browser, the application comprising a sensor including sensor JavaScript code, wherein running the application comprises executing the sensor JavaScript code as the first JavaScript code in the web browser to activate the sensor; and wherein the sensor is configured to: gather data with respect to the application at runtime; and check predetermined application-specific rules against the gathered data for detecting client-side attacks at runtime.

Client-side attack detection in web applications

A sequence of events by a single user with at least one computing system are monitored. Each event characterizes user interaction with the at least one computing system and the sequence of events form a plurality of pairwise disjoint log samples. Thereafter, it is determined, using an adjacency graph trained using a plurality of log samples generated by a plurality of users, whether any of the log samples is anomalous. Data can be provided that characterizes the log samples determined to be anomalous. Related apparatus, systems, techniques and articles are also described.

User classification by local to global sequence alignment techniques for anomaly-based intrusion detection

Various embodiments of systems, computer program products, and methods to automatically generate low-interaction honeypots to protect application landscapes through are described herein. In an aspect, representative applications associated with resources in a network are identified. The low-interaction honeypots are automatically generated for the identified representative applications. Further, the representative applications are probed to retrieve responses corresponding to different requests. Templates are generated corresponding to request-response pairs by parsing the responses and the requests. During operation, new requests for accessing the resources are responded based on the generated templates. The new requests and corresponding responses are recorded.

Automatic generation of low-interaction honeypots

Various examples are directed to detecting anomalous modifications to a software component. For example, a computing device may receive, from a version control system, version metadata describing properties of a plurality of commits for the software component. The computing device may generate a plurality of commit clusters based, at least in part, on the properties of the plurality of commits. The computing device may determine a first anomalous commit of the plurality of commits and generate an alert message indicating a first code segment modified by the first commit.

Anomalous commit detection

Systems and methods are provided herein for establishing a protection framework for a component. Identified assets of a component requiring protection from a potential attack are received. A list of assets is generated based on the identified assets. A protection framework is configured to include at least one defensive pattern to protect the list of assets against the potential attack. The protection framework is executed to establish a hardened boundary between the component and an attack surface of the component.

Component protection frameworks using defensive patterns

Methods, systems, and computer-readable storage media for optimizing query processing in encrypted databases. In some implementations, actions include receiving a query that is to be used to query an encrypted database, generating a plurality of query plans based on the query, each query plan including a local query and one or more remote queries, the local query being executable at a client-side and the one or more remote queries being executable at a server-side, selecting an optimal query plan from the plurality of query plans, providing one or more remote queries of the optimal query plan to the server-side for execution, receiving one or more remote results, and processing a local query of the optimal query plan and the one or more remote results to provide a final query result.

Local versus remote optimization in encrypted query processing

Growing community of inventors

Wiesloch, Germany

Martin Haerterich